operators.py 81.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# function:
#    operators to process sample,
#    eg: decode/resize/crop image

from __future__ import absolute_import
from __future__ import print_function
from __future__ import division

23 24 25 26 27 28 29
try:
    from collections.abc import Sequence
except Exception:
    from collections import Sequence

from numbers import Number

30 31 32 33 34
import uuid
import logging
import random
import math
import numpy as np
35

36 37 38 39
import cv2
from PIL import Image, ImageEnhance

from ppdet.core.workspace import serializable
40
from ppdet.modeling.ops import AnchorGrid
41 42

from .op_helper import (satisfy_sample_constraint, filter_and_process,
43 44
                        generate_sample_bbox, clip_bbox, data_anchor_sampling,
                        satisfy_sample_constraint_coverage, crop_image_sampling,
45
                        generate_sample_bbox_square, bbox_area_sampling,
W
wangguanzhong 已提交
46
                        is_poly, gaussian_radius, draw_gaussian)
47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96

logger = logging.getLogger(__name__)

registered_ops = []


def register_op(cls):
    registered_ops.append(cls.__name__)
    if not hasattr(BaseOperator, cls.__name__):
        setattr(BaseOperator, cls.__name__, cls)
    else:
        raise KeyError("The {} class has been registered.".format(cls.__name__))
    return serializable(cls)


class BboxError(ValueError):
    pass


class ImageError(ValueError):
    pass


class BaseOperator(object):
    def __init__(self, name=None):
        if name is None:
            name = self.__class__.__name__
        self._id = name + '_' + str(uuid.uuid4())[-6:]

    def __call__(self, sample, context=None):
        """ Process a sample.
        Args:
            sample (dict): a dict of sample, eg: {'image':xx, 'label': xxx}
            context (dict): info about this sample processing
        Returns:
            result (dict): a processed sample
        """
        return sample

    def __str__(self):
        return str(self._id)


@register_op
class DecodeImage(BaseOperator):
    def __init__(self, to_rgb=True, with_mixup=False):
        """ Transform the image data to numpy format.

        Args:
            to_rgb (bool): whether to convert BGR to RGB
97
            with_mixup (bool): whether or not to mixup image and gt_bbbox/gt_score
98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
        """

        super(DecodeImage, self).__init__()
        self.to_rgb = to_rgb
        self.with_mixup = with_mixup
        if not isinstance(self.to_rgb, bool):
            raise TypeError("{}: input type is invalid.".format(self))
        if not isinstance(self.with_mixup, bool):
            raise TypeError("{}: input type is invalid.".format(self))

    def __call__(self, sample, context=None):
        """ load image if 'im_file' field is not empty but 'image' is"""
        if 'image' not in sample:
            with open(sample['im_file'], 'rb') as f:
                sample['image'] = f.read()

        im = sample['image']
        data = np.frombuffer(im, dtype='uint8')
        im = cv2.imdecode(data, 1)  # BGR mode, but need RGB mode
        if self.to_rgb:
            im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
        sample['image'] = im

        if 'h' not in sample:
            sample['h'] = im.shape[0]
        if 'w' not in sample:
            sample['w'] = im.shape[1]
        # make default im_info with [h, w, 1]
        sample['im_info'] = np.array(
            [im.shape[0], im.shape[1], 1.], dtype=np.float32)
        # decode mixup image
        if self.with_mixup and 'mixup' in sample:
            self.__call__(sample['mixup'], context)
        return sample


W
wangguanzhong 已提交
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
@register_op
class MultiscaleTestResize(BaseOperator):
    def __init__(self,
                 origin_target_size=800,
                 origin_max_size=1333,
                 target_size=[],
                 max_size=2000,
                 interp=cv2.INTER_LINEAR,
                 use_flip=True):
        """
        Rescale image to the each size in target size, and capped at max_size.

        Args:
            origin_target_size(int): original target size of image's short side.
            origin_max_size(int): original max size of image.
            target_size (list): A list of target sizes of image's short side.
            max_size (int): the max size of image.
            interp (int): the interpolation method.
            use_flip (bool): whether use flip augmentation.
        """
        super(MultiscaleTestResize, self).__init__()
        self.origin_target_size = int(origin_target_size)
        self.origin_max_size = int(origin_max_size)
        self.max_size = int(max_size)
        self.interp = int(interp)
        self.use_flip = use_flip

        if not isinstance(target_size, list):
            raise TypeError(
                "Type of target_size is invalid. Must be List, now is {}".
                format(type(target_size)))
        self.target_size = target_size
        if not (isinstance(self.origin_target_size, int) and isinstance(
                self.origin_max_size, int) and isinstance(self.max_size, int)
                and isinstance(self.interp, int)):
            raise TypeError("{}: input type is invalid.".format(self))

    def __call__(self, sample, context=None):
        """ Resize the image numpy for multi-scale test.
        """
        origin_ims = {}
        im = sample['image']
        if not isinstance(im, np.ndarray):
            raise TypeError("{}: image type is not numpy.".format(self))
        if len(im.shape) != 3:
            raise ImageError('{}: image is not 3-dimensional.'.format(self))
        im_shape = im.shape
        im_size_min = np.min(im_shape[0:2])
        im_size_max = np.max(im_shape[0:2])
        if float(im_size_min) == 0:
            raise ZeroDivisionError('{}: min size of image is 0'.format(self))
        base_name_list = ['image']
        origin_ims['image'] = im
        if self.use_flip:
188 189 190 191
            sample['image_flip'] = im[:, ::-1, :]
            base_name_list.append('image_flip')
            origin_ims['image_flip'] = sample['image_flip']

W
wangguanzhong 已提交
192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208
        for base_name in base_name_list:
            im_scale = float(self.origin_target_size) / float(im_size_min)
            # Prevent the biggest axis from being more than max_size
            if np.round(im_scale * im_size_max) > self.origin_max_size:
                im_scale = float(self.origin_max_size) / float(im_size_max)
            im_scale_x = im_scale
            im_scale_y = im_scale

            resize_w = np.round(im_scale_x * float(im_shape[1]))
            resize_h = np.round(im_scale_y * float(im_shape[0]))
            im_resize = cv2.resize(
                origin_ims[base_name],
                None,
                None,
                fx=im_scale_x,
                fy=im_scale_y,
                interpolation=self.interp)
209

W
wangguanzhong 已提交
210
            sample[base_name] = im_resize
211 212 213 214
            info_name = 'im_info' if base_name == 'image' else 'im_info_image_flip'
            sample[base_name] = im_resize
            sample[info_name] = np.array(
                [resize_h, resize_w, im_scale], dtype=np.float32)
W
wangguanzhong 已提交
215 216 217 218 219 220 221 222 223 224 225 226 227 228 229
            for i, size in enumerate(self.target_size):
                im_scale = float(size) / float(im_size_min)
                if np.round(im_scale * im_size_max) > self.max_size:
                    im_scale = float(self.max_size) / float(im_size_max)
                im_scale_x = im_scale
                im_scale_y = im_scale
                resize_w = np.round(im_scale_x * float(im_shape[1]))
                resize_h = np.round(im_scale_y * float(im_shape[0]))
                im_resize = cv2.resize(
                    origin_ims[base_name],
                    None,
                    None,
                    fx=im_scale_x,
                    fy=im_scale_y,
                    interpolation=self.interp)
230 231 232 233

                im_info = [resize_h, resize_w, im_scale]
                # hard-code here, must be consistent with
                # ppdet/modeling/architectures/input_helper.py
W
wangguanzhong 已提交
234
                name = base_name + '_scale_' + str(i)
235
                info_name = 'im_info_' + name
W
wangguanzhong 已提交
236
                sample[name] = im_resize
237 238
                sample[info_name] = np.array(
                    [resize_h, resize_w, im_scale], dtype=np.float32)
W
wangguanzhong 已提交
239 240 241
        return sample


242 243 244 245 246 247 248 249
@register_op
class ResizeImage(BaseOperator):
    def __init__(self,
                 target_size=0,
                 max_size=0,
                 interp=cv2.INTER_LINEAR,
                 use_cv2=True):
        """
250 251 252 253 254
        Rescale image to the specified target size, and capped at max_size
        if max_size != 0.
        If target_size is list, selected a scale randomly as the specified
        target size.

255
        Args:
256
            target_size (int|list): the target size of image's short side,
257
                multi-scale training is adopted when type is list.
258 259
            max_size (int): the max size of image
            interp (int): the interpolation method
260
            use_cv2 (bool): use the cv2 interpolation method or use PIL
261
                interpolation method
262 263 264 265 266
        """
        super(ResizeImage, self).__init__()
        self.max_size = int(max_size)
        self.interp = int(interp)
        self.use_cv2 = use_cv2
267 268 269 270 271 272 273
        if not (isinstance(target_size, int) or isinstance(target_size, list)):
            raise TypeError(
                "Type of target_size is invalid. Must be Integer or List, now is {}".
                format(type(target_size)))
        self.target_size = target_size
        if not (isinstance(self.max_size, int) and isinstance(self.interp,
                                                              int)):
274 275 276
            raise TypeError("{}: input type is invalid.".format(self))

    def __call__(self, sample, context=None):
W
wangguanzhong 已提交
277
        """ Resize the image numpy.
278 279 280 281 282 283 284 285 286
        """
        im = sample['image']
        if not isinstance(im, np.ndarray):
            raise TypeError("{}: image type is not numpy.".format(self))
        if len(im.shape) != 3:
            raise ImageError('{}: image is not 3-dimensional.'.format(self))
        im_shape = im.shape
        im_size_min = np.min(im_shape[0:2])
        im_size_max = np.max(im_shape[0:2])
287 288 289 290 291
        if isinstance(self.target_size, list):
            # Case for multi-scale training
            selected_size = random.choice(self.target_size)
        else:
            selected_size = self.target_size
292 293 294
        if float(im_size_min) == 0:
            raise ZeroDivisionError('{}: min size of image is 0'.format(self))
        if self.max_size != 0:
295
            im_scale = float(selected_size) / float(im_size_min)
296 297 298 299 300
            # Prevent the biggest axis from being more than max_size
            if np.round(im_scale * im_size_max) > self.max_size:
                im_scale = float(self.max_size) / float(im_size_max)
            im_scale_x = im_scale
            im_scale_y = im_scale
301

W
wangguanzhong 已提交
302 303
            resize_w = im_scale_x * float(im_shape[1])
            resize_h = im_scale_y * float(im_shape[0])
W
wangguanzhong 已提交
304 305 306 307 308 309
            im_info = [resize_h, resize_w, im_scale]
            if 'im_info' in sample and sample['im_info'][2] != 1.:
                sample['im_info'] = np.append(
                    list(sample['im_info']), im_info).astype(np.float32)
            else:
                sample['im_info'] = np.array(im_info).astype(np.float32)
310
        else:
311 312
            im_scale_x = float(selected_size) / float(im_shape[1])
            im_scale_y = float(selected_size) / float(im_shape[0])
313 314 315 316

            resize_w = selected_size
            resize_h = selected_size

317 318 319 320 321 322 323 324 325
        if self.use_cv2:
            im = cv2.resize(
                im,
                None,
                None,
                fx=im_scale_x,
                fy=im_scale_y,
                interpolation=self.interp)
        else:
W
wangguanzhong 已提交
326 327 328 329
            if self.max_size != 0:
                raise TypeError(
                    'If you set max_size to cap the maximum size of image,'
                    'please set use_cv2 to True to resize the image.')
330
            im = im.astype('uint8')
331
            im = Image.fromarray(im)
W
wangguanzhong 已提交
332
            im = im.resize((int(resize_w), int(resize_h)), self.interp)
333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364
            im = np.array(im)

        sample['image'] = im
        return sample


@register_op
class RandomFlipImage(BaseOperator):
    def __init__(self, prob=0.5, is_normalized=False, is_mask_flip=False):
        """
        Args:
            prob (float): the probability of flipping image
            is_normalized (bool): whether the bbox scale to [0,1]
            is_mask_flip (bool): whether flip the segmentation
        """
        super(RandomFlipImage, self).__init__()
        self.prob = prob
        self.is_normalized = is_normalized
        self.is_mask_flip = is_mask_flip
        if not (isinstance(self.prob, float) and
                isinstance(self.is_normalized, bool) and
                isinstance(self.is_mask_flip, bool)):
            raise TypeError("{}: input type is invalid.".format(self))

    def flip_segms(self, segms, height, width):
        def _flip_poly(poly, width):
            flipped_poly = np.array(poly)
            flipped_poly[0::2] = width - np.array(poly[0::2]) - 1
            return flipped_poly.tolist()

        def _flip_rle(rle, height, width):
            if 'counts' in rle and type(rle['counts']) == list:
365
                rle = mask_util.frPyObjects(rle, height, width)
366
            mask = mask_util.decode(rle)
367
            mask = mask[:, ::-1]
368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393
            rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
            return rle

        flipped_segms = []
        for segm in segms:
            if is_poly(segm):
                # Polygon format
                flipped_segms.append([_flip_poly(poly, width) for poly in segm])
            else:
                # RLE format
                import pycocotools.mask as mask_util
                flipped_segms.append(_flip_rle(segm, height, width))
        return flipped_segms

    def __call__(self, sample, context=None):
        """Filp the image and bounding box.
        Operators:
            1. Flip the image numpy.
            2. Transform the bboxes' x coordinates.
              (Must judge whether the coordinates are normalized!)
            3. Transform the segmentations' x coordinates.
              (Must judge whether the coordinates are normalized!)
        Output:
            sample: the image, bounding box and segmentation part
                    in sample are flipped.
        """
394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431

        samples = sample
        batch_input = True
        if not isinstance(samples, Sequence):
            batch_input = False
            samples = [samples]
        for sample in samples:
            gt_bbox = sample['gt_bbox']
            im = sample['image']
            if not isinstance(im, np.ndarray):
                raise TypeError("{}: image is not a numpy array.".format(self))
            if len(im.shape) != 3:
                raise ImageError("{}: image is not 3-dimensional.".format(self))
            height, width, _ = im.shape
            if np.random.uniform(0, 1) < self.prob:
                im = im[:, ::-1, :]
                if gt_bbox.shape[0] == 0:
                    return sample
                oldx1 = gt_bbox[:, 0].copy()
                oldx2 = gt_bbox[:, 2].copy()
                if self.is_normalized:
                    gt_bbox[:, 0] = 1 - oldx2
                    gt_bbox[:, 2] = 1 - oldx1
                else:
                    gt_bbox[:, 0] = width - oldx2 - 1
                    gt_bbox[:, 2] = width - oldx1 - 1
                if gt_bbox.shape[0] != 0 and (
                        gt_bbox[:, 2] < gt_bbox[:, 0]).all():
                    m = "{}: invalid box, x2 should be greater than x1".format(
                        self)
                    raise BboxError(m)
                sample['gt_bbox'] = gt_bbox
                if self.is_mask_flip and len(sample['gt_poly']) != 0:
                    sample['gt_poly'] = self.flip_segms(sample['gt_poly'],
                                                        height, width)
                sample['flipped'] = True
                sample['image'] = im
        sample = samples if batch_input else samples[0]
432 433 434
        return sample


littletomatodonkey's avatar
littletomatodonkey 已提交
435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503
@register_op
class AutoAugmentImage(BaseOperator):
    def __init__(self, is_normalized=False, autoaug_type="v1"):
        """
        Args:
            is_normalized (bool): whether the bbox scale to [0,1]
            autoaug_type (str): autoaug type, support v0, v1, v2, v3, test
        """
        super(AutoAugmentImage, self).__init__()
        self.is_normalized = is_normalized
        self.autoaug_type = autoaug_type
        if not isinstance(self.is_normalized, bool):
            raise TypeError("{}: input type is invalid.".format(self))

    def __call__(self, sample, context=None):
        """
        Learning Data Augmentation Strategies for Object Detection, see https://arxiv.org/abs/1906.11172
        """
        samples = sample
        batch_input = True
        if not isinstance(samples, Sequence):
            batch_input = False
            samples = [samples]
        for sample in samples:
            gt_bbox = sample['gt_bbox']
            im = sample['image']
            if not isinstance(im, np.ndarray):
                raise TypeError("{}: image is not a numpy array.".format(self))
            if len(im.shape) != 3:
                raise ImageError("{}: image is not 3-dimensional.".format(self))
            if len(gt_bbox) == 0:
                continue

            # gt_boxes : [x1, y1, x2, y2]
            # norm_gt_boxes: [y1, x1, y2, x2]
            height, width, _ = im.shape
            norm_gt_bbox = np.ones_like(gt_bbox, dtype=np.float32)
            if not self.is_normalized:
                norm_gt_bbox[:, 0] = gt_bbox[:, 1] / float(height)
                norm_gt_bbox[:, 1] = gt_bbox[:, 0] / float(width)
                norm_gt_bbox[:, 2] = gt_bbox[:, 3] / float(height)
                norm_gt_bbox[:, 3] = gt_bbox[:, 2] / float(width)
            else:
                norm_gt_bbox[:, 0] = gt_bbox[:, 1]
                norm_gt_bbox[:, 1] = gt_bbox[:, 0]
                norm_gt_bbox[:, 2] = gt_bbox[:, 3]
                norm_gt_bbox[:, 3] = gt_bbox[:, 2]

            from .autoaugment_utils import distort_image_with_autoaugment
            im, norm_gt_bbox = distort_image_with_autoaugment(im, norm_gt_bbox,
                                                              self.autoaug_type)
            if not self.is_normalized:
                gt_bbox[:, 0] = norm_gt_bbox[:, 1] * float(width)
                gt_bbox[:, 1] = norm_gt_bbox[:, 0] * float(height)
                gt_bbox[:, 2] = norm_gt_bbox[:, 3] * float(width)
                gt_bbox[:, 3] = norm_gt_bbox[:, 2] * float(height)
            else:
                gt_bbox[:, 0] = norm_gt_bbox[:, 1]
                gt_bbox[:, 1] = norm_gt_bbox[:, 0]
                gt_bbox[:, 2] = norm_gt_bbox[:, 3]
                gt_bbox[:, 3] = norm_gt_bbox[:, 2]

            sample['gt_bbox'] = gt_bbox
            sample['image'] = im

        sample = samples if batch_input else samples[0]
        return sample


504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533
@register_op
class NormalizeImage(BaseOperator):
    def __init__(self,
                 mean=[0.485, 0.456, 0.406],
                 std=[1, 1, 1],
                 is_scale=True,
                 is_channel_first=True):
        """
        Args:
            mean (list): the pixel mean
            std (list): the pixel variance
        """
        super(NormalizeImage, self).__init__()
        self.mean = mean
        self.std = std
        self.is_scale = is_scale
        self.is_channel_first = is_channel_first
        if not (isinstance(self.mean, list) and isinstance(self.std, list) and
                isinstance(self.is_scale, bool)):
            raise TypeError("{}: input type is invalid.".format(self))
        from functools import reduce
        if reduce(lambda x, y: x * y, self.std) == 0:
            raise ValueError('{}: std is invalid!'.format(self))

    def __call__(self, sample, context=None):
        """Normalize the image.
        Operators:
            1.(optional) Scale the image to [0,1]
            2. Each pixel minus mean and is divided by std
        """
534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558
        samples = sample
        batch_input = True
        if not isinstance(samples, Sequence):
            batch_input = False
            samples = [samples]
        for sample in samples:
            for k in sample.keys():
                # hard code
                if k.startswith('image'):
                    im = sample[k]
                    im = im.astype(np.float32, copy=False)
                    if self.is_channel_first:
                        mean = np.array(self.mean)[:, np.newaxis, np.newaxis]
                        std = np.array(self.std)[:, np.newaxis, np.newaxis]
                    else:
                        mean = np.array(self.mean)[np.newaxis, np.newaxis, :]
                        std = np.array(self.std)[np.newaxis, np.newaxis, :]
                    if self.is_scale:
                        im = im / 255.0
                    im -= mean
                    im /= std
                    sample[k] = im
        if not batch_input:
            samples = samples[0]
        return samples
559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675


@register_op
class RandomDistort(BaseOperator):
    def __init__(self,
                 brightness_lower=0.5,
                 brightness_upper=1.5,
                 contrast_lower=0.5,
                 contrast_upper=1.5,
                 saturation_lower=0.5,
                 saturation_upper=1.5,
                 hue_lower=-18,
                 hue_upper=18,
                 brightness_prob=0.5,
                 contrast_prob=0.5,
                 saturation_prob=0.5,
                 hue_prob=0.5,
                 count=4,
                 is_order=False):
        """
        Args:
            brightness_lower/ brightness_upper (float): the brightness
                between brightness_lower and brightness_upper
            contrast_lower/ contrast_upper (float): the contrast between
                contrast_lower and contrast_lower
            saturation_lower/ saturation_upper (float): the saturation
                between saturation_lower and saturation_upper
            hue_lower/ hue_upper (float): the hue between
                hue_lower and hue_upper
            brightness_prob (float): the probability of changing brightness
            contrast_prob (float): the probability of changing contrast
            saturation_prob (float): the probability of changing saturation
            hue_prob (float): the probability of changing hue
            count (int): the kinds of doing distrot
            is_order (bool): whether determine the order of distortion
        """
        super(RandomDistort, self).__init__()
        self.brightness_lower = brightness_lower
        self.brightness_upper = brightness_upper
        self.contrast_lower = contrast_lower
        self.contrast_upper = contrast_upper
        self.saturation_lower = saturation_lower
        self.saturation_upper = saturation_upper
        self.hue_lower = hue_lower
        self.hue_upper = hue_upper
        self.brightness_prob = brightness_prob
        self.contrast_prob = contrast_prob
        self.saturation_prob = saturation_prob
        self.hue_prob = hue_prob
        self.count = count
        self.is_order = is_order

    def random_brightness(self, img):
        brightness_delta = np.random.uniform(self.brightness_lower,
                                             self.brightness_upper)
        prob = np.random.uniform(0, 1)
        if prob < self.brightness_prob:
            img = ImageEnhance.Brightness(img).enhance(brightness_delta)
        return img

    def random_contrast(self, img):
        contrast_delta = np.random.uniform(self.contrast_lower,
                                           self.contrast_upper)
        prob = np.random.uniform(0, 1)
        if prob < self.contrast_prob:
            img = ImageEnhance.Contrast(img).enhance(contrast_delta)
        return img

    def random_saturation(self, img):
        saturation_delta = np.random.uniform(self.saturation_lower,
                                             self.saturation_upper)
        prob = np.random.uniform(0, 1)
        if prob < self.saturation_prob:
            img = ImageEnhance.Color(img).enhance(saturation_delta)
        return img

    def random_hue(self, img):
        hue_delta = np.random.uniform(self.hue_lower, self.hue_upper)
        prob = np.random.uniform(0, 1)
        if prob < self.hue_prob:
            img = np.array(img.convert('HSV'))
            img[:, :, 0] = img[:, :, 0] + hue_delta
            img = Image.fromarray(img, mode='HSV').convert('RGB')
        return img

    def __call__(self, sample, context):
        """random distort the image"""
        ops = [
            self.random_brightness, self.random_contrast,
            self.random_saturation, self.random_hue
        ]
        if self.is_order:
            prob = np.random.uniform(0, 1)
            if prob < 0.5:
                ops = [
                    self.random_brightness,
                    self.random_saturation,
                    self.random_hue,
                    self.random_contrast,
                ]
        else:
            ops = random.sample(ops, self.count)
        assert 'image' in sample, "image data not found"
        im = sample['image']
        im = Image.fromarray(im)
        for id in range(self.count):
            im = ops[id](im)
        im = np.asarray(im)
        sample['image'] = im
        return sample


@register_op
class ExpandImage(BaseOperator):
    def __init__(self, max_ratio, prob, mean=[127.5, 127.5, 127.5]):
        """
        Args:
676
            max_ratio (float): the ratio of expanding
677 678 679 680 681 682 683 684 685 686 687 688
            prob (float): the probability of expanding image
            mean (list): the pixel mean
        """
        super(ExpandImage, self).__init__()
        self.max_ratio = max_ratio
        self.mean = mean
        self.prob = prob

    def __call__(self, sample, context):
        """
        Expand the image and modify bounding box.
        Operators:
689
            1. Scale the image width and height.
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751
            2. Construct new images with new height and width.
            3. Fill the new image with the mean.
            4. Put original imge into new image.
            5. Rescale the bounding box.
            6. Determine if the new bbox is satisfied in the new image.
        Returns:
            sample: the image, bounding box are replaced.
        """

        prob = np.random.uniform(0, 1)
        assert 'image' in sample, 'not found image data'
        im = sample['image']
        gt_bbox = sample['gt_bbox']
        gt_class = sample['gt_class']
        im_width = sample['w']
        im_height = sample['h']
        if prob < self.prob:
            if self.max_ratio - 1 >= 0.01:
                expand_ratio = np.random.uniform(1, self.max_ratio)
                height = int(im_height * expand_ratio)
                width = int(im_width * expand_ratio)
                h_off = math.floor(np.random.uniform(0, height - im_height))
                w_off = math.floor(np.random.uniform(0, width - im_width))
                expand_bbox = [
                    -w_off / im_width, -h_off / im_height,
                    (width - w_off) / im_width, (height - h_off) / im_height
                ]
                expand_im = np.ones((height, width, 3))
                expand_im = np.uint8(expand_im * np.squeeze(self.mean))
                expand_im = Image.fromarray(expand_im)
                im = Image.fromarray(im)
                expand_im.paste(im, (int(w_off), int(h_off)))
                expand_im = np.asarray(expand_im)
                gt_bbox, gt_class, _ = filter_and_process(expand_bbox, gt_bbox,
                                                          gt_class)
                sample['image'] = expand_im
                sample['gt_bbox'] = gt_bbox
                sample['gt_class'] = gt_class
                sample['w'] = width
                sample['h'] = height

        return sample


@register_op
class CropImage(BaseOperator):
    def __init__(self, batch_sampler, satisfy_all=False, avoid_no_bbox=True):
        """
        Args:
            batch_sampler (list): Multiple sets of different
                                  parameters for cropping.
            satisfy_all (bool): whether all boxes must satisfy.
            e.g.[[1, 1, 1.0, 1.0, 1.0, 1.0, 0.0, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.1, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.3, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.5, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.7, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.9, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.0, 1.0]]
           [max sample, max trial, min scale, max scale,
            min aspect ratio, max aspect ratio,
            min overlap, max overlap]
752
            avoid_no_bbox (bool): whether to to avoid the
753
                                  situation where the box does not appear.
754 755 756 757 758 759 760 761 762 763
        """
        super(CropImage, self).__init__()
        self.batch_sampler = batch_sampler
        self.satisfy_all = satisfy_all
        self.avoid_no_bbox = avoid_no_bbox

    def __call__(self, sample, context):
        """
        Crop the image and modify bounding box.
        Operators:
764
            1. Scale the image width and height.
765 766 767 768 769 770 771 772 773 774 775 776
            2. Crop the image according to a radom sample.
            3. Rescale the bounding box.
            4. Determine if the new bbox is satisfied in the new image.
        Returns:
            sample: the image, bounding box are replaced.
        """
        assert 'image' in sample, "image data not found"
        im = sample['image']
        gt_bbox = sample['gt_bbox']
        gt_class = sample['gt_class']
        im_width = sample['w']
        im_height = sample['h']
777 778 779
        gt_score = None
        if 'gt_score' in sample:
            gt_score = sample['gt_score']
780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814
        sampled_bbox = []
        gt_bbox = gt_bbox.tolist()
        for sampler in self.batch_sampler:
            found = 0
            for i in range(sampler[1]):
                if found >= sampler[0]:
                    break
                sample_bbox = generate_sample_bbox(sampler)
                if satisfy_sample_constraint(sampler, sample_bbox, gt_bbox,
                                             self.satisfy_all):
                    sampled_bbox.append(sample_bbox)
                    found = found + 1
        im = np.array(im)
        while sampled_bbox:
            idx = int(np.random.uniform(0, len(sampled_bbox)))
            sample_bbox = sampled_bbox.pop(idx)
            sample_bbox = clip_bbox(sample_bbox)
            crop_bbox, crop_class, crop_score = \
                filter_and_process(sample_bbox, gt_bbox, gt_class, gt_score)
            if self.avoid_no_bbox:
                if len(crop_bbox) < 1:
                    continue
            xmin = int(sample_bbox[0] * im_width)
            xmax = int(sample_bbox[2] * im_width)
            ymin = int(sample_bbox[1] * im_height)
            ymax = int(sample_bbox[3] * im_height)
            im = im[ymin:ymax, xmin:xmax]
            sample['image'] = im
            sample['gt_bbox'] = crop_bbox
            sample['gt_class'] = crop_class
            sample['gt_score'] = crop_score
            return sample
        return sample


815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853
@register_op
class CropImageWithDataAchorSampling(BaseOperator):
    def __init__(self,
                 batch_sampler,
                 anchor_sampler=None,
                 target_size=None,
                 das_anchor_scales=[16, 32, 64, 128],
                 sampling_prob=0.5,
                 min_size=8.,
                 avoid_no_bbox=True):
        """
        Args:
            anchor_sampler (list): anchor_sampling sets of different
                                  parameters for cropping.
            batch_sampler (list): Multiple sets of different
                                  parameters for cropping.
              e.g.[[1, 10, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.2, 0.0]]
                  [[1, 50, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0],
                   [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0],
                   [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0],
                   [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0],
                   [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0]]
              [max sample, max trial, min scale, max scale,
               min aspect ratio, max aspect ratio,
               min overlap, max overlap, min coverage, max coverage]
            target_size (bool): target image size.
            das_anchor_scales (list[float]): a list of anchor scales in data
                anchor smapling.
            min_size (float): minimum size of sampled bbox.
            avoid_no_bbox (bool): whether to to avoid the
                                  situation where the box does not appear.
        """
        super(CropImageWithDataAchorSampling, self).__init__()
        self.anchor_sampler = anchor_sampler
        self.batch_sampler = batch_sampler
        self.target_size = target_size
        self.sampling_prob = sampling_prob
        self.min_size = min_size
        self.avoid_no_bbox = avoid_no_bbox
854
        self.das_anchor_scales = np.array(das_anchor_scales)
855 856 857 858 859

    def __call__(self, sample, context):
        """
        Crop the image and modify bounding box.
        Operators:
860
            1. Scale the image width and height.
861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887
            2. Crop the image according to a radom sample.
            3. Rescale the bounding box.
            4. Determine if the new bbox is satisfied in the new image.
        Returns:
            sample: the image, bounding box are replaced.
        """
        assert 'image' in sample, "image data not found"
        im = sample['image']
        gt_bbox = sample['gt_bbox']
        gt_class = sample['gt_class']
        image_width = sample['w']
        image_height = sample['h']
        gt_score = None
        if 'gt_score' in sample:
            gt_score = sample['gt_score']
        sampled_bbox = []
        gt_bbox = gt_bbox.tolist()

        prob = np.random.uniform(0., 1.)
        if prob > self.sampling_prob:  # anchor sampling
            assert self.anchor_sampler
            for sampler in self.anchor_sampler:
                found = 0
                for i in range(sampler[1]):
                    if found >= sampler[0]:
                        break
                    sample_bbox = data_anchor_sampling(
888 889
                        gt_bbox, image_width, image_height,
                        self.das_anchor_scales, self.target_size)
890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959
                    if sample_bbox == 0:
                        break
                    if satisfy_sample_constraint_coverage(sampler, sample_bbox,
                                                          gt_bbox):
                        sampled_bbox.append(sample_bbox)
                        found = found + 1
            im = np.array(im)
            while sampled_bbox:
                idx = int(np.random.uniform(0, len(sampled_bbox)))
                sample_bbox = sampled_bbox.pop(idx)

                crop_bbox, crop_class, crop_score = filter_and_process(
                    sample_bbox, gt_bbox, gt_class, gt_score)
                crop_bbox, crop_class, crop_score = bbox_area_sampling(
                    crop_bbox, crop_class, crop_score, self.target_size,
                    self.min_size)

                if self.avoid_no_bbox:
                    if len(crop_bbox) < 1:
                        continue
                im = crop_image_sampling(im, sample_bbox, image_width,
                                         image_height, self.target_size)
                sample['image'] = im
                sample['gt_bbox'] = crop_bbox
                sample['gt_class'] = crop_class
                sample['gt_score'] = crop_score
                return sample
            return sample

        else:
            for sampler in self.batch_sampler:
                found = 0
                for i in range(sampler[1]):
                    if found >= sampler[0]:
                        break
                    sample_bbox = generate_sample_bbox_square(
                        sampler, image_width, image_height)
                    if satisfy_sample_constraint_coverage(sampler, sample_bbox,
                                                          gt_bbox):
                        sampled_bbox.append(sample_bbox)
                        found = found + 1
            im = np.array(im)
            while sampled_bbox:
                idx = int(np.random.uniform(0, len(sampled_bbox)))
                sample_bbox = sampled_bbox.pop(idx)
                sample_bbox = clip_bbox(sample_bbox)

                crop_bbox, crop_class, crop_score = filter_and_process(
                    sample_bbox, gt_bbox, gt_class, gt_score)
                # sampling bbox according the bbox area
                crop_bbox, crop_class, crop_score = bbox_area_sampling(
                    crop_bbox, crop_class, crop_score, self.target_size,
                    self.min_size)

                if self.avoid_no_bbox:
                    if len(crop_bbox) < 1:
                        continue
                xmin = int(sample_bbox[0] * image_width)
                xmax = int(sample_bbox[2] * image_width)
                ymin = int(sample_bbox[1] * image_height)
                ymax = int(sample_bbox[3] * image_height)
                im = im[ymin:ymax, xmin:xmax]
                sample['image'] = im
                sample['gt_bbox'] = crop_bbox
                sample['gt_class'] = crop_class
                sample['gt_score'] = crop_score
                return sample
            return sample


960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997
@register_op
class NormalizeBox(BaseOperator):
    """Transform the bounding box's coornidates to [0,1]."""

    def __init__(self):
        super(NormalizeBox, self).__init__()

    def __call__(self, sample, context):
        gt_bbox = sample['gt_bbox']
        width = sample['w']
        height = sample['h']
        for i in range(gt_bbox.shape[0]):
            gt_bbox[i][0] = gt_bbox[i][0] / width
            gt_bbox[i][1] = gt_bbox[i][1] / height
            gt_bbox[i][2] = gt_bbox[i][2] / width
            gt_bbox[i][3] = gt_bbox[i][3] / height
        sample['gt_bbox'] = gt_bbox
        return sample


@register_op
class Permute(BaseOperator):
    def __init__(self, to_bgr=True, channel_first=True):
        """
        Change the channel.
        Args:
            to_bgr (bool): confirm whether to convert RGB to BGR
            channel_first (bool): confirm whether to change channel

        """
        super(Permute, self).__init__()
        self.to_bgr = to_bgr
        self.channel_first = channel_first
        if not (isinstance(self.to_bgr, bool) and
                isinstance(self.channel_first, bool)):
            raise TypeError("{}: input type is invalid.".format(self))

    def __call__(self, sample, context=None):
998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017
        samples = sample
        batch_input = True
        if not isinstance(samples, Sequence):
            batch_input = False
            samples = [samples]
        for sample in samples:
            assert 'image' in sample, "image data not found"
            for k in sample.keys():
                # hard code
                if k.startswith('image'):
                    im = sample[k]
                    if self.channel_first:
                        im = np.swapaxes(im, 1, 2)
                        im = np.swapaxes(im, 1, 0)
                    if self.to_bgr:
                        im = im[[2, 1, 0], :, :]
                    sample[k] = im
        if not batch_input:
            samples = samples[0]
        return samples
1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067


@register_op
class MixupImage(BaseOperator):
    def __init__(self, alpha=1.5, beta=1.5):
        """ Mixup image and gt_bbbox/gt_score
        Args:
            alpha (float): alpha parameter of beta distribute
            beta (float): beta parameter of beta distribute
        """
        super(MixupImage, self).__init__()
        self.alpha = alpha
        self.beta = beta
        if self.alpha <= 0.0:
            raise ValueError("alpha shold be positive in {}".format(self))
        if self.beta <= 0.0:
            raise ValueError("beta shold be positive in {}".format(self))

    def _mixup_img(self, img1, img2, factor):
        h = max(img1.shape[0], img2.shape[0])
        w = max(img1.shape[1], img2.shape[1])
        img = np.zeros((h, w, img1.shape[2]), 'float32')
        img[:img1.shape[0], :img1.shape[1], :] = \
            img1.astype('float32') * factor
        img[:img2.shape[0], :img2.shape[1], :] += \
            img2.astype('float32') * (1.0 - factor)
        return img.astype('uint8')

    def __call__(self, sample, context=None):
        if 'mixup' not in sample:
            return sample
        factor = np.random.beta(self.alpha, self.beta)
        factor = max(0.0, min(1.0, factor))
        if factor >= 1.0:
            sample.pop('mixup')
            return sample
        if factor <= 0.0:
            return sample['mixup']
        im = self._mixup_img(sample['image'], sample['mixup']['image'], factor)
        gt_bbox1 = sample['gt_bbox']
        gt_bbox2 = sample['mixup']['gt_bbox']
        gt_bbox = np.concatenate((gt_bbox1, gt_bbox2), axis=0)
        gt_class1 = sample['gt_class']
        gt_class2 = sample['mixup']['gt_class']
        gt_class = np.concatenate((gt_class1, gt_class2), axis=0)

        gt_score1 = sample['gt_score']
        gt_score2 = sample['mixup']['gt_score']
        gt_score = np.concatenate(
            (gt_score1 * factor, gt_score2 * (1. - factor)), axis=0)
1068 1069 1070 1071 1072

        is_crowd1 = sample['is_crowd']
        is_crowd2 = sample['mixup']['is_crowd']
        is_crowd = np.concatenate((is_crowd1, is_crowd2), axis=0)

1073 1074 1075 1076
        sample['image'] = im
        sample['gt_bbox'] = gt_bbox
        sample['gt_score'] = gt_score
        sample['gt_class'] = gt_class
1077
        sample['is_crowd'] = is_crowd
1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113
        sample['h'] = im.shape[0]
        sample['w'] = im.shape[1]
        sample.pop('mixup')
        return sample


@register_op
class RandomInterpImage(BaseOperator):
    def __init__(self, target_size=0, max_size=0):
        """
        Random reisze image by multiply interpolate method.
        Args:
            target_size (int): the taregt size of image's short side
            max_size (int): the max size of image
        """
        super(RandomInterpImage, self).__init__()
        self.target_size = target_size
        self.max_size = max_size
        if not (isinstance(self.target_size, int) and
                isinstance(self.max_size, int)):
            raise TypeError('{}: input type is invalid.'.format(self))
        interps = [
            cv2.INTER_NEAREST,
            cv2.INTER_LINEAR,
            cv2.INTER_AREA,
            cv2.INTER_CUBIC,
            cv2.INTER_LANCZOS4,
        ]
        self.resizers = []
        for interp in interps:
            self.resizers.append(ResizeImage(target_size, max_size, interp))

    def __call__(self, sample, context=None):
        """Resise the image numpy by random resizer."""
        resizer = random.choice(self.resizers)
        return resizer(sample, context)
1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126


@register_op
class Resize(BaseOperator):
    """Resize image and bbox.

    Args:
        target_dim (int or list): target size, can be a single number or a list
            (for random shape).
        interp (int or str): interpolation method, can be an integer or
            'random' (for randomized interpolation).
            default to `cv2.INTER_LINEAR`.
    """
W
wangguanzhong 已提交
1127 1128

    def __init__(self, target_dim=[], interp=cv2.INTER_LINEAR):
1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148
        super(Resize, self).__init__()
        self.target_dim = target_dim
        self.interp = interp  # 'random' for yolov3

    def __call__(self, sample, context=None):
        w = sample['w']
        h = sample['h']

        interp = self.interp
        if interp == 'random':
            interp = np.random.choice(range(5))

        if isinstance(self.target_dim, Sequence):
            dim = np.random.choice(self.target_dim)
        else:
            dim = self.target_dim
        resize_w = resize_h = dim
        scale_x = dim / w
        scale_y = dim / h
        if 'gt_bbox' in sample and len(sample['gt_bbox']) > 0:
W
wangguanzhong 已提交
1149 1150 1151
            scale_array = np.array([scale_x, scale_y] * 2, dtype=np.float32)
            sample['gt_bbox'] = np.clip(sample['gt_bbox'] * scale_array, 0,
                                        dim - 1)
1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175
        sample['h'] = resize_h
        sample['w'] = resize_w

        sample['image'] = cv2.resize(
            sample['image'], (resize_w, resize_h), interpolation=interp)
        return sample


@register_op
class ColorDistort(BaseOperator):
    """Random color distortion.

    Args:
        hue (list): hue settings.
            in [lower, upper, probability] format.
        saturation (list): saturation settings.
            in [lower, upper, probability] format.
        contrast (list): contrast settings.
            in [lower, upper, probability] format.
        brightness (list): brightness settings.
            in [lower, upper, probability] format.
        random_apply (bool): whether to apply in random (yolo) or fixed (SSD)
            order.
    """
W
wangguanzhong 已提交
1176

1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200
    def __init__(self,
                 hue=[-18, 18, 0.5],
                 saturation=[0.5, 1.5, 0.5],
                 contrast=[0.5, 1.5, 0.5],
                 brightness=[0.5, 1.5, 0.5],
                 random_apply=True):
        super(ColorDistort, self).__init__()
        self.hue = hue
        self.saturation = saturation
        self.contrast = contrast
        self.brightness = brightness
        self.random_apply = random_apply

    def apply_hue(self, img):
        low, high, prob = self.hue
        if np.random.uniform(0., 1.) < prob:
            return img

        img = img.astype(np.float32)

        # XXX works, but result differ from HSV version
        delta = np.random.uniform(low, high)
        u = np.cos(delta * np.pi)
        w = np.sin(delta * np.pi)
W
wangguanzhong 已提交
1201 1202
        bt = np.array([[1.0, 0.0, 0.0], [0.0, u, -w], [0.0, w, u]])
        tyiq = np.array([[0.299, 0.587, 0.114], [0.596, -0.274, -0.321],
1203
                         [0.211, -0.523, 0.311]])
W
wangguanzhong 已提交
1204
        ityiq = np.array([[1.0, 0.956, 0.621], [1.0, -0.272, -0.647],
1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246
                          [1.0, -1.107, 1.705]])
        t = np.dot(np.dot(ityiq, bt), tyiq).T
        img = np.dot(img, t)
        return img

    def apply_saturation(self, img):
        low, high, prob = self.saturation
        if np.random.uniform(0., 1.) < prob:
            return img
        delta = np.random.uniform(low, high)

        img = img.astype(np.float32)
        gray = img * np.array([[[0.299, 0.587, 0.114]]], dtype=np.float32)
        gray = gray.sum(axis=2, keepdims=True)
        gray *= (1.0 - delta)
        img *= delta
        img += gray
        return img

    def apply_contrast(self, img):
        low, high, prob = self.contrast
        if np.random.uniform(0., 1.) < prob:
            return img
        delta = np.random.uniform(low, high)

        img = img.astype(np.float32)
        img *= delta
        return img

    def apply_brightness(self, img):
        low, high, prob = self.brightness
        if np.random.uniform(0., 1.) < prob:
            return img
        delta = np.random.uniform(low, high)

        img = img.astype(np.float32)
        img += delta
        return img

    def __call__(self, sample, context=None):
        img = sample['image']
        if self.random_apply:
W
wangguanzhong 已提交
1247 1248 1249 1250 1251 1252 1253
            functions = [
                self.apply_brightness,
                self.apply_contrast,
                self.apply_saturation,
                self.apply_hue,
            ]
            distortions = np.random.permutation(functions)
1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272
            for func in distortions:
                img = func(img)
            sample['image'] = img
            return sample

        img = self.apply_brightness(img)

        if np.random.randint(0, 2):
            img = self.apply_contrast(img)
            img = self.apply_saturation(img)
            img = self.apply_hue(img)
        else:
            img = self.apply_saturation(img)
            img = self.apply_hue(img)
            img = self.apply_contrast(img)
        sample['image'] = img
        return sample


W
wangguanzhong 已提交
1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332
@register_op
class CornerRandColor(ColorDistort):
    """Random color for CornerNet series models.
    Args:
        saturation (float): saturation settings.
        contrast (float): contrast settings.
        brightness (float): brightness settings.
        is_scale (bool): whether to scale the input image.
    """

    def __init__(self,
                 saturation=0.4,
                 contrast=0.4,
                 brightness=0.4,
                 is_scale=True):
        super(CornerRandColor, self).__init__(
            saturation=saturation, contrast=contrast, brightness=brightness)
        self.is_scale = is_scale

    def apply_saturation(self, img, img_gray):
        alpha = 1. + np.random.uniform(
            low=-self.saturation, high=self.saturation)
        self._blend(alpha, img, img_gray[:, :, None])
        return img

    def apply_contrast(self, img, img_gray):
        alpha = 1. + np.random.uniform(low=-self.contrast, high=self.contrast)
        img_mean = img_gray.mean()
        self._blend(alpha, img, img_mean)
        return img

    def apply_brightness(self, img, img_gray):
        alpha = 1 + np.random.uniform(
            low=-self.brightness, high=self.brightness)
        img *= alpha
        return img

    def _blend(self, alpha, img, img_mean):
        img *= alpha
        img_mean *= (1 - alpha)
        img += img_mean

    def __call__(self, sample, context=None):
        img = sample['image']
        if self.is_scale:
            img = img.astype(np.float32, copy=False)
            img /= 255.
        img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
        functions = [
            self.apply_brightness,
            self.apply_contrast,
            self.apply_saturation,
        ]
        distortions = np.random.permutation(functions)
        for func in distortions:
            img = func(img, img_gray)
        sample['image'] = img
        return sample


1333 1334 1335 1336 1337 1338 1339 1340
@register_op
class NormalizePermute(BaseOperator):
    """Normalize and permute channel order.

    Args:
        mean (list): mean values in RGB order.
        std (list): std values in RGB order.
    """
W
wangguanzhong 已提交
1341

1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370
    def __init__(self,
                 mean=[123.675, 116.28, 103.53],
                 std=[58.395, 57.120, 57.375]):
        super(NormalizePermute, self).__init__()
        self.mean = mean
        self.std = std

    def __call__(self, sample, context=None):
        img = sample['image']
        img = img.astype(np.float32)

        img = img.transpose((2, 0, 1))
        mean = np.array(self.mean, dtype=np.float32)
        std = np.array(self.std, dtype=np.float32)
        invstd = 1. / std
        for v, m, s in zip(img, mean, invstd):
            v.__isub__(m).__imul__(s)
        sample['image'] = img
        return sample


@register_op
class RandomExpand(BaseOperator):
    """Random expand the canvas.

    Args:
        ratio (float): maximum expansion ratio.
        prob (float): probability to expand.
        fill_value (list): color value used to fill the canvas. in RGB order.
1371
        is_mask_expand(bool): whether expand the segmentation.
1372
    """
W
wangguanzhong 已提交
1373

1374 1375 1376 1377 1378
    def __init__(self,
                 ratio=4.,
                 prob=0.5,
                 fill_value=(127.5, ) * 3,
                 is_mask_expand=False):
1379 1380 1381 1382 1383 1384 1385
        super(RandomExpand, self).__init__()
        assert ratio > 1.01, "expand ratio must be larger than 1.01"
        self.ratio = ratio
        self.prob = prob
        assert isinstance(fill_value, (Number, Sequence)), \
            "fill value must be either float or sequence"
        if isinstance(fill_value, Number):
W
wangguanzhong 已提交
1386
            fill_value = (fill_value, ) * 3
1387 1388 1389
        if not isinstance(fill_value, tuple):
            fill_value = tuple(fill_value)
        self.fill_value = fill_value
1390
        self.is_mask_expand = is_mask_expand
1391

1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423
    def expand_segms(self, segms, x, y, height, width, ratio):
        def _expand_poly(poly, x, y):
            expanded_poly = np.array(poly)
            expanded_poly[0::2] += x
            expanded_poly[1::2] += y
            return expanded_poly.tolist()

        def _expand_rle(rle, x, y, height, width, ratio):
            if 'counts' in rle and type(rle['counts']) == list:
                rle = mask_util.frPyObjects(rle, height, width)
            mask = mask_util.decode(rle)
            expanded_mask = np.full((int(height * ratio), int(width * ratio)),
                                    0).astype(mask.dtype)
            expanded_mask[y:y + height, x:x + width] = mask
            rle = mask_util.encode(
                np.array(
                    expanded_mask, order='F', dtype=np.uint8))
            return rle

        expanded_segms = []
        for segm in segms:
            if is_poly(segm):
                # Polygon format
                expanded_segms.append(
                    [_expand_poly(poly, x, y) for poly in segm])
            else:
                # RLE format
                import pycocotools.mask as mask_util
                expanded_segms.append(
                    _expand_rle(segm, x, y, height, width, ratio))
        return expanded_segms

1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447
    def __call__(self, sample, context=None):
        if np.random.uniform(0., 1.) < self.prob:
            return sample

        img = sample['image']
        height = int(sample['h'])
        width = int(sample['w'])

        expand_ratio = np.random.uniform(1., self.ratio)
        h = int(height * expand_ratio)
        w = int(width * expand_ratio)
        if not h > height or not w > width:
            return sample
        y = np.random.randint(0, h - height)
        x = np.random.randint(0, w - width)
        canvas = np.ones((h, w, 3), dtype=np.uint8)
        canvas *= np.array(self.fill_value, dtype=np.uint8)
        canvas[y:y + height, x:x + width, :] = img.astype(np.uint8)

        sample['h'] = h
        sample['w'] = w
        sample['image'] = canvas
        if 'gt_bbox' in sample and len(sample['gt_bbox']) > 0:
            sample['gt_bbox'] += np.array([x, y] * 2, dtype=np.float32)
1448 1449
        if self.is_mask_expand and 'gt_poly' in sample and len(sample[
                'gt_poly']) > 0:
1450 1451
            sample['gt_poly'] = self.expand_segms(sample['gt_poly'], x, y,
                                                  height, width, expand_ratio)
1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467
        return sample


@register_op
class RandomCrop(BaseOperator):
    """Random crop image and bboxes.

    Args:
        aspect_ratio (list): aspect ratio of cropped region.
            in [min, max] format.
        thresholds (list): iou thresholds for decide a valid bbox crop.
        scaling (list): ratio between a cropped region and the original image.
             in [min, max] format.
        num_attempts (int): number of tries before giving up.
        allow_no_crop (bool): allow return without actually cropping them.
        cover_all_box (bool): ensure all bboxes are covered in the final crop.
1468
        is_mask_crop(bool): whether crop the segmentation.
1469
    """
W
wangguanzhong 已提交
1470

1471 1472 1473 1474 1475 1476
    def __init__(self,
                 aspect_ratio=[.5, 2.],
                 thresholds=[.0, .1, .3, .5, .7, .9],
                 scaling=[.3, 1.],
                 num_attempts=50,
                 allow_no_crop=True,
1477 1478
                 cover_all_box=False,
                 is_mask_crop=False):
1479 1480 1481 1482 1483 1484 1485
        super(RandomCrop, self).__init__()
        self.aspect_ratio = aspect_ratio
        self.thresholds = thresholds
        self.scaling = scaling
        self.num_attempts = num_attempts
        self.allow_no_crop = allow_no_crop
        self.cover_all_box = cover_all_box
1486
        self.is_mask_crop = is_mask_crop
1487

1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556
    def crop_segms(self, segms, valid_ids, crop, height, width):
        def _crop_poly(segm, crop):
            xmin, ymin, xmax, ymax = crop
            crop_coord = [xmin, ymin, xmin, ymax, xmax, ymax, xmax, ymin]
            crop_p = np.array(crop_coord).reshape(4, 2)
            crop_p = Polygon(crop_p)

            crop_segm = list()
            for poly in segm:
                poly = np.array(poly).reshape(len(poly) // 2, 2)
                polygon = Polygon(poly)
                if not polygon.is_valid:
                    exterior = polygon.exterior
                    multi_lines = exterior.intersection(exterior)
                    polygons = shapely.ops.polygonize(multi_lines)
                    polygon = MultiPolygon(polygons)
                multi_polygon = list()
                if isinstance(polygon, MultiPolygon):
                    multi_polygon = copy.deepcopy(polygon)
                else:
                    multi_polygon.append(copy.deepcopy(polygon))
                for per_polygon in multi_polygon:
                    inter = per_polygon.intersection(crop_p)
                    if not inter:
                        continue
                    if isinstance(inter, (MultiPolygon, GeometryCollection)):
                        for part in inter:
                            if not isinstance(part, Polygon):
                                continue
                            part = np.squeeze(
                                np.array(part.exterior.coords[:-1]).reshape(1,
                                                                            -1))
                            part[0::2] -= xmin
                            part[1::2] -= ymin
                            crop_segm.append(part.tolist())
                    elif isinstance(inter, Polygon):
                        crop_poly = np.squeeze(
                            np.array(inter.exterior.coords[:-1]).reshape(1, -1))
                        crop_poly[0::2] -= xmin
                        crop_poly[1::2] -= ymin
                        crop_segm.append(crop_poly.tolist())
                    else:
                        continue
            return crop_segm

        def _crop_rle(rle, crop, height, width):
            if 'counts' in rle and type(rle['counts']) == list:
                rle = mask_util.frPyObjects(rle, height, width)
            mask = mask_util.decode(rle)
            mask = mask[crop[1]:crop[3], crop[0]:crop[2]]
            rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
            return rle

        crop_segms = []
        for id in valid_ids:
            segm = segms[id]
            if is_poly(segm):
                import copy
                import shapely.ops
                from shapely.geometry import Polygon, MultiPolygon, GeometryCollection
                logging.getLogger("shapely").setLevel(logging.WARNING)
                # Polygon format
                crop_segms.append(_crop_poly(segm, crop))
            else:
                # RLE format
                import pycocotools.mask as mask_util
                crop_segms.append(_crop_rle(segm, crop, height, width))
        return crop_segms

1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585
    def __call__(self, sample, context=None):
        if 'gt_bbox' in sample and len(sample['gt_bbox']) == 0:
            return sample

        h = sample['h']
        w = sample['w']
        gt_bbox = sample['gt_bbox']

        # NOTE Original method attempts to generate one candidate for each
        # threshold then randomly sample one from the resulting list.
        # Here a short circuit approach is taken, i.e., randomly choose a
        # threshold and attempt to find a valid crop, and simply return the
        # first one found.
        # The probability is not exactly the same, kinda resembling the
        # "Monty Hall" problem. Actually carrying out the attempts will affect
        # observability (just like opening doors in the "Monty Hall" game).
        thresholds = list(self.thresholds)
        if self.allow_no_crop:
            thresholds.append('no_crop')
        np.random.shuffle(thresholds)

        for thresh in thresholds:
            if thresh == 'no_crop':
                return sample

            found = False
            for i in range(self.num_attempts):
                scale = np.random.uniform(*self.scaling)
                min_ar, max_ar = self.aspect_ratio
W
wangguanzhong 已提交
1586 1587
                aspect_ratio = np.random.uniform(
                    max(min_ar, scale**2), min(max_ar, scale**-2))
1588 1589 1590 1591 1592
                crop_h = int(h * scale / np.sqrt(aspect_ratio))
                crop_w = int(w * scale * np.sqrt(aspect_ratio))
                crop_y = np.random.randint(0, h - crop_h)
                crop_x = np.random.randint(0, w - crop_w)
                crop_box = [crop_x, crop_y, crop_x + crop_w, crop_y + crop_h]
W
wangguanzhong 已提交
1593 1594 1595
                iou = self._iou_matrix(
                    gt_bbox, np.array(
                        [crop_box], dtype=np.float32))
1596 1597 1598 1599 1600 1601 1602
                if iou.max() < thresh:
                    continue

                if self.cover_all_box and iou.min() < thresh:
                    continue

                cropped_box, valid_ids = self._crop_box_with_center_constraint(
W
wangguanzhong 已提交
1603 1604
                    gt_bbox, np.array(
                        crop_box, dtype=np.float32))
1605 1606 1607 1608 1609
                if valid_ids.size > 0:
                    found = True
                    break

            if found:
1610 1611
                if self.is_mask_crop and 'gt_poly' in sample and len(sample[
                        'gt_poly']) > 0:
1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632
                    crop_polys = self.crop_segms(
                        sample['gt_poly'],
                        valid_ids,
                        np.array(
                            crop_box, dtype=np.int64),
                        h,
                        w)
                    if [] in crop_polys:
                        delete_id = list()
                        valid_polys = list()
                        for id, crop_poly in enumerate(crop_polys):
                            if crop_poly == []:
                                delete_id.append(id)
                            else:
                                valid_polys.append(crop_poly)
                        valid_ids = np.delete(valid_ids, delete_id)
                        if len(valid_polys) == 0:
                            return sample
                        sample['gt_poly'] = valid_polys
                    else:
                        sample['gt_poly'] = crop_polys
1633 1634 1635 1636 1637 1638 1639 1640 1641
                sample['image'] = self._crop_image(sample['image'], crop_box)
                sample['gt_bbox'] = np.take(cropped_box, valid_ids, axis=0)
                sample['gt_class'] = np.take(
                    sample['gt_class'], valid_ids, axis=0)
                sample['w'] = crop_box[2] - crop_box[0]
                sample['h'] = crop_box[3] - crop_box[1]
                if 'gt_score' in sample:
                    sample['gt_score'] = np.take(
                        sample['gt_score'], valid_ids, axis=0)
1642 1643 1644 1645

                if 'is_crowd' in sample:
                    sample['is_crowd'] = np.take(
                        sample['is_crowd'], valid_ids, axis=0)
1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668
                return sample

        return sample

    def _iou_matrix(self, a, b):
        tl_i = np.maximum(a[:, np.newaxis, :2], b[:, :2])
        br_i = np.minimum(a[:, np.newaxis, 2:], b[:, 2:])

        area_i = np.prod(br_i - tl_i, axis=2) * (tl_i < br_i).all(axis=2)
        area_a = np.prod(a[:, 2:] - a[:, :2], axis=1)
        area_b = np.prod(b[:, 2:] - b[:, :2], axis=1)
        area_o = (area_a[:, np.newaxis] + area_b - area_i)
        return area_i / (area_o + 1e-10)

    def _crop_box_with_center_constraint(self, box, crop):
        cropped_box = box.copy()

        cropped_box[:, :2] = np.maximum(box[:, :2], crop[:2])
        cropped_box[:, 2:] = np.minimum(box[:, 2:], crop[2:])
        cropped_box[:, :2] -= crop[:2]
        cropped_box[:, 2:] -= crop[:2]

        centers = (box[:, :2] + box[:, 2:]) / 2
W
wangguanzhong 已提交
1669 1670
        valid = np.logical_and(crop[:2] <= centers,
                               centers < crop[2:]).all(axis=1)
1671 1672 1673 1674 1675 1676 1677 1678
        valid = np.logical_and(
            valid, (cropped_box[:, :2] < cropped_box[:, 2:]).all(axis=1))

        return cropped_box, np.where(valid)[0]

    def _crop_image(self, img, crop):
        x1, y1, x2, y2 = crop
        return img[y1:y2, x1:x2, :]
1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738


@register_op
class PadBox(BaseOperator):
    def __init__(self, num_max_boxes=50):
        """
        Pad zeros to bboxes if number of bboxes is less than num_max_boxes.
        Args:
            num_max_boxes (int): the max number of bboxes
        """
        self.num_max_boxes = num_max_boxes
        super(PadBox, self).__init__()

    def __call__(self, sample, context=None):
        assert 'gt_bbox' in sample
        bbox = sample['gt_bbox']
        gt_num = min(self.num_max_boxes, len(bbox))
        num_max = self.num_max_boxes
        fields = context['fields'] if context else []
        pad_bbox = np.zeros((num_max, 4), dtype=np.float32)
        if gt_num > 0:
            pad_bbox[:gt_num, :] = bbox[:gt_num, :]
        sample['gt_bbox'] = pad_bbox
        if 'gt_class' in fields:
            pad_class = np.zeros((num_max), dtype=np.int32)
            if gt_num > 0:
                pad_class[:gt_num] = sample['gt_class'][:gt_num, 0]
            sample['gt_class'] = pad_class
        if 'gt_score' in fields:
            pad_score = np.zeros((num_max), dtype=np.float32)
            if gt_num > 0:
                pad_score[:gt_num] = sample['gt_score'][:gt_num, 0]
            sample['gt_score'] = pad_score
        # in training, for example in op ExpandImage,
        # the bbox and gt_class is expandded, but the difficult is not,
        # so, judging by it's length
        if 'is_difficult' in fields:
            pad_diff = np.zeros((num_max), dtype=np.int32)
            if gt_num > 0:
                pad_diff[:gt_num] = sample['difficult'][:gt_num, 0]
            sample['difficult'] = pad_diff
        return sample


@register_op
class BboxXYXY2XYWH(BaseOperator):
    """
    Convert bbox XYXY format to XYWH format.
    """

    def __init__(self):
        super(BboxXYXY2XYWH, self).__init__()

    def __call__(self, sample, context=None):
        assert 'gt_bbox' in sample
        bbox = sample['gt_bbox']
        bbox[:, 2:4] = bbox[:, 2:4] - bbox[:, :2]
        bbox[:, :2] = bbox[:, :2] + bbox[:, 2:4] / 2.
        sample['gt_bbox'] = bbox
        return sample
W
wangguanzhong 已提交
1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974


@register_op
class Lighting(BaseOperator):
    """
    Lighting the imagen by eigenvalues and eigenvectors
    Args:
        eigval (list): eigenvalues
        eigvec (list): eigenvectors
        alphastd (float): random weight of lighting, 0.1 by default
    """

    def __init__(self, eigval, eigvec, alphastd=0.1):
        super(Lighting, self).__init__()
        self.alphastd = alphastd
        self.eigval = np.array(eigval).astype('float32')
        self.eigvec = np.array(eigvec).astype('float32')

    def __call__(self, sample, context=None):
        alpha = np.random.normal(scale=self.alphastd, size=(3, ))
        sample['image'] += np.dot(self.eigvec, self.eigval * alpha)
        return sample


@register_op
class CornerTarget(BaseOperator):
    """
    Generate targets for CornerNet by ground truth data. 
    Args:
        output_size (int): the size of output heatmaps.
        num_classes (int): num of classes.
        gaussian_bump (bool): whether to apply gaussian bump on gt targets.
            True by default.
        gaussian_rad (int): radius of gaussian bump. If it is set to -1, the 
            radius will be calculated by iou. -1 by default.
        gaussian_iou (float): the threshold iou of predicted bbox to gt bbox. 
            If the iou is larger than threshold, the predicted bboox seems as
            positive sample. 0.3 by default
        max_tag_len (int): max num of gt box per image.
    """

    def __init__(self,
                 output_size,
                 num_classes,
                 gaussian_bump=True,
                 gaussian_rad=-1,
                 gaussian_iou=0.3,
                 max_tag_len=128):
        super(CornerTarget, self).__init__()
        self.num_classes = num_classes
        self.output_size = output_size
        self.gaussian_bump = gaussian_bump
        self.gaussian_rad = gaussian_rad
        self.gaussian_iou = gaussian_iou
        self.max_tag_len = max_tag_len

    def __call__(self, sample, context=None):
        tl_heatmaps = np.zeros(
            (self.num_classes, self.output_size[0], self.output_size[1]),
            dtype=np.float32)
        br_heatmaps = np.zeros(
            (self.num_classes, self.output_size[0], self.output_size[1]),
            dtype=np.float32)

        tl_regrs = np.zeros((self.max_tag_len, 2), dtype=np.float32)
        br_regrs = np.zeros((self.max_tag_len, 2), dtype=np.float32)
        tl_tags = np.zeros((self.max_tag_len), dtype=np.int64)
        br_tags = np.zeros((self.max_tag_len), dtype=np.int64)
        tag_masks = np.zeros((self.max_tag_len), dtype=np.uint8)
        tag_lens = np.zeros((), dtype=np.int32)
        tag_nums = np.zeros((1), dtype=np.int32)

        gt_bbox = sample['gt_bbox']
        gt_class = sample['gt_class']
        keep_inds  = ((gt_bbox[:, 2] - gt_bbox[:, 0]) > 0) & \
                ((gt_bbox[:, 3] - gt_bbox[:, 1]) > 0)
        gt_bbox = gt_bbox[keep_inds]
        gt_class = gt_class[keep_inds]
        sample['gt_bbox'] = gt_bbox
        sample['gt_class'] = gt_class
        width_ratio = self.output_size[1] / sample['w']
        height_ratio = self.output_size[0] / sample['h']
        for i in range(gt_bbox.shape[0]):
            width = gt_bbox[i][2] - gt_bbox[i][0]
            height = gt_bbox[i][3] - gt_bbox[i][1]

            xtl, ytl = gt_bbox[i][0], gt_bbox[i][1]
            xbr, ybr = gt_bbox[i][2], gt_bbox[i][3]

            fxtl = (xtl * width_ratio)
            fytl = (ytl * height_ratio)
            fxbr = (xbr * width_ratio)
            fybr = (ybr * height_ratio)

            xtl = int(fxtl)
            ytl = int(fytl)
            xbr = int(fxbr)
            ybr = int(fybr)
            if self.gaussian_bump:
                width = math.ceil(width * width_ratio)
                height = math.ceil(height * height_ratio)
                if self.gaussian_rad == -1:
                    radius = gaussian_radius((height, width), self.gaussian_iou)
                    radius = max(0, int(radius))
                else:
                    radius = self.gaussian_rad
                draw_gaussian(tl_heatmaps[gt_class[i][0]], [xtl, ytl], radius)
                draw_gaussian(br_heatmaps[gt_class[i][0]], [xbr, ybr], radius)
            else:
                tl_heatmaps[gt_class[i][0], ytl, xtl] = 1
                br_heatmaps[gt_class[i][0], ybr, xbr] = 1

            tl_regrs[i, :] = [fxtl - xtl, fytl - ytl]
            br_regrs[i, :] = [fxbr - xbr, fybr - ybr]
            tl_tags[tag_lens] = ytl * self.output_size[1] + xtl
            br_tags[tag_lens] = ybr * self.output_size[1] + xbr
            tag_lens += 1

        tag_masks[:tag_lens] = 1

        sample['tl_heatmaps'] = tl_heatmaps
        sample['br_heatmaps'] = br_heatmaps
        sample['tl_regrs'] = tl_regrs
        sample['br_regrs'] = br_regrs
        sample['tl_tags'] = tl_tags
        sample['br_tags'] = br_tags
        sample['tag_masks'] = tag_masks

        return sample


@register_op
class CornerCrop(BaseOperator):
    """
    Random crop for CornerNet
    Args:
        random_scales (list): scales of output_size to input_size.
        border (int): border of corp center
        is_train (bool): train or test
        input_size (int): size of input image
    """

    def __init__(self,
                 random_scales=[0.6, 0.7, 0.8, 0.9, 1., 1.1, 1.2, 1.3],
                 border=128,
                 is_train=True,
                 input_size=511):
        super(CornerCrop, self).__init__()
        self.random_scales = random_scales
        self.border = border
        self.is_train = is_train
        self.input_size = input_size

    def __call__(self, sample, context=None):
        im_h, im_w = int(sample['h']), int(sample['w'])
        if self.is_train:
            scale = np.random.choice(self.random_scales)
            height = int(self.input_size * scale)
            width = int(self.input_size * scale)

            w_border = self._get_border(self.border, im_w)
            h_border = self._get_border(self.border, im_h)

            ctx = np.random.randint(low=w_border, high=im_w - w_border)
            cty = np.random.randint(low=h_border, high=im_h - h_border)

        else:
            cty, ctx = im_h // 2, im_w // 2
            height = im_h | 127
            width = im_w | 127

        cropped_image = np.zeros(
            (height, width, 3), dtype=sample['image'].dtype)

        x0, x1 = max(ctx - width // 2, 0), min(ctx + width // 2, im_w)
        y0, y1 = max(cty - height // 2, 0), min(cty + height // 2, im_h)

        left_w, right_w = ctx - x0, x1 - ctx
        top_h, bottom_h = cty - y0, y1 - cty

        # crop image
        cropped_ctx, cropped_cty = width // 2, height // 2
        x_slice = slice(int(cropped_ctx - left_w), int(cropped_ctx + right_w))
        y_slice = slice(int(cropped_cty - top_h), int(cropped_cty + bottom_h))
        cropped_image[y_slice, x_slice, :] = sample['image'][y0:y1, x0:x1, :]

        sample['image'] = cropped_image
        sample['h'], sample['w'] = height, width

        if self.is_train:
            # crop detections
            gt_bbox = sample['gt_bbox']
            gt_bbox[:, 0:4:2] -= x0
            gt_bbox[:, 1:4:2] -= y0
            gt_bbox[:, 0:4:2] += cropped_ctx - left_w
            gt_bbox[:, 1:4:2] += cropped_cty - top_h
        else:
            sample['borders'] = np.array(
                [
                    cropped_cty - top_h, cropped_cty + bottom_h,
                    cropped_ctx - left_w, cropped_ctx + right_w
                ],
                dtype=np.float32)

        return sample

    def _get_border(self, border, size):
        i = 1
        while size - border // i <= border // i:
            i *= 2
        return border // i


@register_op
class CornerRatio(BaseOperator):
    """
    Ratio of output size to image size
    Args:
        input_size (int): the size of input size
        output_size (int): the size of heatmap
    """

    def __init__(self, input_size=511, output_size=64):
        super(CornerRatio, self).__init__()
        self.input_size = input_size
        self.output_size = output_size

    def __call__(self, sample, context=None):
        scale = (self.input_size + 1) // self.output_size
        out_height, out_width = (sample['h'] + 1) // scale, (
            sample['w'] + 1) // scale
        height_ratio = out_height / float(sample['h'])
        width_ratio = out_width / float(sample['w'])
        sample['ratios'] = np.array([height_ratio, width_ratio])

        return sample
1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173


@register_op
class RandomScaledCrop(BaseOperator):
    """Resize image and bbox based on long side (with optional random scaling),
       then crop or pad image to target size.

    Args:
        target_dim (int): target size.
        scale_range (list): random scale range.
        interp (int): interpolation method, default to `cv2.INTER_LINEAR`.
    """

    def __init__(self,
                 target_dim=512,
                 scale_range=[.1, 2.],
                 interp=cv2.INTER_LINEAR):
        super(RandomScaledCrop, self).__init__()
        self.target_dim = target_dim
        self.scale_range = scale_range
        self.interp = interp

    def __call__(self, sample, context=None):
        w = sample['w']
        h = sample['h']
        random_scale = np.random.uniform(*self.scale_range)
        dim = self.target_dim
        random_dim = int(dim * random_scale)
        dim_max = max(h, w)
        scale = random_dim / dim_max
        resize_w = int(round(w * scale))
        resize_h = int(round(h * scale))
        offset_x = int(max(0, np.random.uniform(0., resize_w - dim)))
        offset_y = int(max(0, np.random.uniform(0., resize_h - dim)))
        if 'gt_bbox' in sample and len(sample['gt_bbox']) > 0:
            scale_array = np.array([scale, scale] * 2, dtype=np.float32)
            shift_array = np.array([offset_x, offset_y] * 2, dtype=np.float32)
            boxes = sample['gt_bbox'] * scale_array - shift_array
            boxes = np.clip(boxes, 0, dim - 1)
            # filter boxes with no area
            area = np.prod(boxes[..., 2:] - boxes[..., :2], axis=1)
            valid = (area > 1.).nonzero()[0]
            sample['gt_bbox'] = boxes[valid]
            sample['gt_class'] = sample['gt_class'][valid]

        img = sample['image']
        img = cv2.resize(img, (resize_w, resize_h), interpolation=self.interp)
        img = np.array(img)
        canvas = np.zeros((dim, dim, 3), dtype=img.dtype)
        canvas[:min(dim, resize_h), :min(dim, resize_w), :] = img[
            offset_y:offset_y + dim, offset_x:offset_x + dim, :]
        sample['h'] = dim
        sample['w'] = dim
        sample['image'] = canvas
        sample['im_info'] = [resize_h, resize_w, scale]
        return sample


@register_op
class ResizeAndPad(BaseOperator):
    """Resize image and bbox, then pad image to target size.

    Args:
        target_dim (int): target size
        interp (int): interpolation method, default to `cv2.INTER_LINEAR`.
    """

    def __init__(self, target_dim=512, interp=cv2.INTER_LINEAR):
        super(ResizeAndPad, self).__init__()
        self.target_dim = target_dim
        self.interp = interp

    def __call__(self, sample, context=None):
        w = sample['w']
        h = sample['h']
        interp = self.interp
        dim = self.target_dim
        dim_max = max(h, w)
        scale = self.target_dim / dim_max
        resize_w = int(round(w * scale))
        resize_h = int(round(h * scale))
        if 'gt_bbox' in sample and len(sample['gt_bbox']) > 0:
            scale_array = np.array([scale, scale] * 2, dtype=np.float32)
            sample['gt_bbox'] = np.clip(sample['gt_bbox'] * scale_array, 0,
                                        dim - 1)
        img = sample['image']
        img = cv2.resize(img, (resize_w, resize_h), interpolation=interp)
        img = np.array(img)
        canvas = np.zeros((dim, dim, 3), dtype=img.dtype)
        canvas[:resize_h, :resize_w, :] = img
        sample['h'] = dim
        sample['w'] = dim
        sample['image'] = canvas
        sample['im_info'] = [resize_h, resize_w, scale]
        return sample


@register_op
class TargetAssign(BaseOperator):
    """Assign regression target and labels.

    Args:
        image_size (int or list): input image size, a single integer or list of
            [h, w]. Default: 512
        min_level (int): min level of the feature pyramid. Default: 3
        max_level (int): max level of the feature pyramid. Default: 7
        anchor_base_scale (int): base anchor scale. Default: 4
        num_scales (int): number of anchor scales. Default: 3
        aspect_ratios (list): aspect ratios.
            Default: [(1, 1), (1.4, 0.7), (0.7, 1.4)]
        match_threshold (float): threshold for foreground IoU. Default: 0.5
    """

    def __init__(self,
                 image_size=512,
                 min_level=3,
                 max_level=7,
                 anchor_base_scale=4,
                 num_scales=3,
                 aspect_ratios=[(1, 1), (1.4, 0.7), (0.7, 1.4)],
                 match_threshold=0.5):
        super(TargetAssign, self).__init__()
        assert image_size % 2 ** max_level == 0, \
            "image size should be multiple of the max level stride"
        self.image_size = image_size
        self.min_level = min_level
        self.max_level = max_level
        self.anchor_base_scale = anchor_base_scale
        self.num_scales = num_scales
        self.aspect_ratios = aspect_ratios
        self.match_threshold = match_threshold

    @property
    def anchors(self):
        if not hasattr(self, '_anchors'):
            anchor_grid = AnchorGrid(self.image_size, self.min_level,
                                     self.max_level, self.anchor_base_scale,
                                     self.num_scales, self.aspect_ratios)
            self._anchors = np.concatenate(anchor_grid.generate())
        return self._anchors

    def iou_matrix(self, a, b):
        tl_i = np.maximum(a[:, np.newaxis, :2], b[:, :2])
        br_i = np.minimum(a[:, np.newaxis, 2:], b[:, 2:])
        area_i = np.prod(br_i - tl_i, axis=2) * (tl_i < br_i).all(axis=2)
        area_a = np.prod(a[:, 2:] - a[:, :2], axis=1)
        area_b = np.prod(b[:, 2:] - b[:, :2], axis=1)
        area_o = (area_a[:, np.newaxis] + area_b - area_i)
        # return area_i / (area_o + 1e-10)
        return np.where(area_i == 0., np.zeros_like(area_i), area_i / area_o)

    def match(self, anchors, gt_boxes):
        # XXX put smaller matrix first would be a little bit faster
        mat = self.iou_matrix(gt_boxes, anchors)
        max_anchor_for_each_gt = mat.argmax(axis=1)
        max_for_each_anchor = mat.max(axis=0)
        anchor_to_gt = mat.argmax(axis=0)
        anchor_to_gt[max_for_each_anchor < self.match_threshold] = -1
        # XXX ensure each gt has at least one anchor assigned,
        # see `force_match_for_each_row` in TF implementation
        one_hot = np.zeros_like(mat)
        one_hot[np.arange(mat.shape[0]), max_anchor_for_each_gt] = 1.
        max_anchor_indices = one_hot.sum(axis=0).nonzero()[0]
        max_gt_indices = one_hot.argmax(axis=0)[max_anchor_indices]
        anchor_to_gt[max_anchor_indices] = max_gt_indices
        return anchor_to_gt

    def encode(self, anchors, boxes):
        wha = anchors[..., 2:] - anchors[..., :2] + 1
        ca = anchors[..., :2] + wha * .5
        whb = boxes[..., 2:] - boxes[..., :2] + 1
        cb = boxes[..., :2] + whb * .5
        offsets = np.empty_like(anchors)
        offsets[..., :2] = (cb - ca) / wha
        offsets[..., 2:] = np.log(whb / wha)
        return offsets

    def __call__(self, sample, context=None):
        gt_boxes = sample['gt_bbox']
        gt_labels = sample['gt_class']
        labels = np.full((self.anchors.shape[0], 1), 0, dtype=np.int32)
        targets = np.full((self.anchors.shape[0], 4), 0., dtype=np.float32)
        sample['gt_label'] = labels
        sample['gt_target'] = targets

        if len(gt_boxes) < 1:
            sample['fg_num'] = np.array(0, dtype=np.int32)
            return sample

        anchor_to_gt = self.match(self.anchors, gt_boxes)
        matched_indices = (anchor_to_gt >= 0).nonzero()[0]
        labels[matched_indices] = gt_labels[anchor_to_gt[matched_indices]]

        matched_boxes = gt_boxes[anchor_to_gt[matched_indices]]
        matched_anchors = self.anchors[matched_indices]
        matched_targets = self.encode(matched_anchors, matched_boxes)
        targets[matched_indices] = matched_targets
        sample['fg_num'] = np.array(len(matched_targets), dtype=np.int32)
        return sample