batch_operators.py 24.3 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 23 24 25 26 27 28
# 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.

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

try:
    from collections.abc import Sequence
except Exception:
    from collections import Sequence

import logging
import cv2
import numpy as np

from .operators import register_op, BaseOperator
W
wangguanzhong 已提交
29
from .op_helper import jaccard_overlap, gaussian2D
30 31 32

logger = logging.getLogger(__name__)

33
__all__ = [
W
wangguanzhong 已提交
34 35 36 37 38 39
    'PadBatch',
    'RandomShape',
    'PadMultiScaleTest',
    'Gt2YoloTarget',
    'Gt2FCOSTarget',
    'Gt2TTFTarget',
40
]
K
Kaipeng Deng 已提交
41

42 43 44 45 46 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

@register_op
class PadBatch(BaseOperator):
    """
    Pad a batch of samples so they can be divisible by a stride.
    The layout of each image should be 'CHW'.
    Args:
        pad_to_stride (int): If `pad_to_stride > 0`, pad zeros to ensure
            height and width is divisible by `pad_to_stride`.
    """

    def __init__(self, pad_to_stride=0, use_padded_im_info=True):
        super(PadBatch, self).__init__()
        self.pad_to_stride = pad_to_stride
        self.use_padded_im_info = use_padded_im_info

    def __call__(self, samples, context=None):
        """
        Args:
            samples (list): a batch of sample, each is dict.
        """
        coarsest_stride = self.pad_to_stride
        if coarsest_stride == 0:
            return samples
        max_shape = np.array([data['image'].shape for data in samples]).max(
            axis=0)

        if coarsest_stride > 0:
            max_shape[1] = int(
                np.ceil(max_shape[1] / coarsest_stride) * coarsest_stride)
            max_shape[2] = int(
                np.ceil(max_shape[2] / coarsest_stride) * coarsest_stride)

        padding_batch = []
        for data in samples:
            im = data['image']
            im_c, im_h, im_w = im.shape[:]
            padding_im = np.zeros(
                (im_c, max_shape[1], max_shape[2]), dtype=np.float32)
            padding_im[:, :im_h, :im_w] = im
            data['image'] = padding_im
            if self.use_padded_im_info:
                data['im_info'][:2] = max_shape[1:3]
S
sunxl1988 已提交
85 86 87 88 89 90 91
            if 'semantic' in data.keys() and data['semantic'] is not None:
                semantic = data['semantic']
                padding_sem = np.zeros(
                    (1, max_shape[1], max_shape[2]), dtype=np.float32)
                padding_sem[:, :im_h, :im_w] = semantic
                data['semantic'] = padding_sem

92 93 94 95 96 97 98 99 100 101 102 103 104 105 106
        return samples


@register_op
class RandomShape(BaseOperator):
    """
    Randomly reshape a batch. If random_inter is True, also randomly
    select one an interpolation algorithm [cv2.INTER_NEAREST, cv2.INTER_LINEAR,
    cv2.INTER_AREA, cv2.INTER_CUBIC, cv2.INTER_LANCZOS4]. If random_inter is
    False, use cv2.INTER_NEAREST.
    Args:
        sizes (list): list of int, random choose a size from these
        random_inter (bool): whether to randomly interpolation, defalut true.
    """

W
wangguanzhong 已提交
107
    def __init__(self, sizes=[], random_inter=False, resize_box=False):
108 109 110 111 112 113 114 115 116 117
        super(RandomShape, self).__init__()
        self.sizes = sizes
        self.random_inter = random_inter
        self.interps = [
            cv2.INTER_NEAREST,
            cv2.INTER_LINEAR,
            cv2.INTER_AREA,
            cv2.INTER_CUBIC,
            cv2.INTER_LANCZOS4,
        ] if random_inter else []
W
wangguanzhong 已提交
118
        self.resize_box = resize_box
119 120 121 122 123 124 125 126 127 128 129 130 131

    def __call__(self, samples, context=None):
        shape = np.random.choice(self.sizes)
        method = np.random.choice(self.interps) if self.random_inter \
            else cv2.INTER_NEAREST
        for i in range(len(samples)):
            im = samples[i]['image']
            h, w = im.shape[:2]
            scale_x = float(shape) / w
            scale_y = float(shape) / h
            im = cv2.resize(
                im, None, None, fx=scale_x, fy=scale_y, interpolation=method)
            samples[i]['image'] = im
W
wangguanzhong 已提交
132 133 134 135 136 137
            if self.resize_box and 'gt_bbox' in samples[i] and len(samples[0][
                    'gt_bbox']) > 0:
                scale_array = np.array([scale_x, scale_y] * 2, dtype=np.float32)
                samples[i]['gt_bbox'] = np.clip(samples[i]['gt_bbox'] *
                                                scale_array, 0,
                                                float(shape) - 1)
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188
        return samples


@register_op
class PadMultiScaleTest(BaseOperator):
    """
    Pad the image so they can be divisible by a stride for multi-scale testing.
 
    Args:
        pad_to_stride (int): If `pad_to_stride > 0`, pad zeros to ensure
            height and width is divisible by `pad_to_stride`.
    """

    def __init__(self, pad_to_stride=0):
        super(PadMultiScaleTest, self).__init__()
        self.pad_to_stride = pad_to_stride

    def __call__(self, samples, context=None):
        coarsest_stride = self.pad_to_stride
        if coarsest_stride == 0:
            return samples

        batch_input = True
        if not isinstance(samples, Sequence):
            batch_input = False
            samples = [samples]
        if len(samples) != 1:
            raise ValueError("Batch size must be 1 when using multiscale test, "
                             "but now batch size is {}".format(len(samples)))
        for i in range(len(samples)):
            sample = samples[i]
            for k in sample.keys():
                # hard code
                if k.startswith('image'):
                    im = sample[k]
                    im_c, im_h, im_w = im.shape
                    max_h = int(
                        np.ceil(im_h / coarsest_stride) * coarsest_stride)
                    max_w = int(
                        np.ceil(im_w / coarsest_stride) * coarsest_stride)
                    padding_im = np.zeros(
                        (im_c, max_h, max_w), dtype=np.float32)

                    padding_im[:, :im_h, :im_w] = im
                    sample[k] = padding_im
                    info_name = 'im_info' if k == 'image' else 'im_info_' + k
                    # update im_info
                    sample[info_name][:2] = [max_h, max_w]
        if not batch_input:
            samples = samples[0]
        return samples
K
Kaipeng Deng 已提交
189 190 191 192 193 194 195 196 197


@register_op
class Gt2YoloTarget(BaseOperator):
    """
    Generate YOLOv3 targets by groud truth data, this operator is only used in
    fine grained YOLOv3 loss mode
    """

W
wangguanzhong 已提交
198 199 200 201 202 203
    def __init__(self,
                 anchors,
                 anchor_masks,
                 downsample_ratios,
                 num_classes=80,
                 iou_thresh=1.):
K
Kaipeng Deng 已提交
204 205 206 207 208
        super(Gt2YoloTarget, self).__init__()
        self.anchors = anchors
        self.anchor_masks = anchor_masks
        self.downsample_ratios = downsample_ratios
        self.num_classes = num_classes
W
wangguanzhong 已提交
209
        self.iou_thresh = iou_thresh
K
Kaipeng Deng 已提交
210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248

    def __call__(self, samples, context=None):
        assert len(self.anchor_masks) == len(self.downsample_ratios), \
            "anchor_masks', and 'downsample_ratios' should have same length."

        h, w = samples[0]['image'].shape[1:3]
        an_hw = np.array(self.anchors) / np.array([[w, h]])
        for sample in samples:
            # im, gt_bbox, gt_class, gt_score = sample
            im = sample['image']
            gt_bbox = sample['gt_bbox']
            gt_class = sample['gt_class']
            gt_score = sample['gt_score']
            for i, (
                    mask, downsample_ratio
            ) in enumerate(zip(self.anchor_masks, self.downsample_ratios)):
                grid_h = int(h / downsample_ratio)
                grid_w = int(w / downsample_ratio)
                target = np.zeros(
                    (len(mask), 6 + self.num_classes, grid_h, grid_w),
                    dtype=np.float32)
                for b in range(gt_bbox.shape[0]):
                    gx, gy, gw, gh = gt_bbox[b, :]
                    cls = gt_class[b]
                    score = gt_score[b]
                    if gw <= 0. or gh <= 0. or score <= 0.:
                        continue

                    # find best match anchor index
                    best_iou = 0.
                    best_idx = -1
                    for an_idx in range(an_hw.shape[0]):
                        iou = jaccard_overlap(
                            [0., 0., gw, gh],
                            [0., 0., an_hw[an_idx, 0], an_hw[an_idx, 1]])
                        if iou > best_iou:
                            best_iou = iou
                            best_idx = an_idx

W
wangguanzhong 已提交
249 250 251
                    gi = int(gx * grid_w)
                    gj = int(gy * grid_h)

K
Kaipeng Deng 已提交
252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270
                    # gtbox should be regresed in this layes if best match 
                    # anchor index in anchor mask of this layer
                    if best_idx in mask:
                        best_n = mask.index(best_idx)

                        # x, y, w, h, scale
                        target[best_n, 0, gj, gi] = gx * grid_w - gi
                        target[best_n, 1, gj, gi] = gy * grid_h - gj
                        target[best_n, 2, gj, gi] = np.log(
                            gw * w / self.anchors[best_idx][0])
                        target[best_n, 3, gj, gi] = np.log(
                            gh * h / self.anchors[best_idx][1])
                        target[best_n, 4, gj, gi] = 2.0 - gw * gh

                        # objectness record gt_score
                        target[best_n, 5, gj, gi] = score

                        # classification
                        target[best_n, 6 + cls, gj, gi] = 1.
W
wangguanzhong 已提交
271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294

                    # For non-matched anchors, calculate the target if the iou 
                    # between anchor and gt is larger than iou_thresh
                    if self.iou_thresh < 1:
                        for idx, mask_i in enumerate(mask):
                            if mask_i == best_idx: continue
                            iou = jaccard_overlap(
                                [0., 0., gw, gh],
                                [0., 0., an_hw[mask_i, 0], an_hw[mask_i, 1]])
                            if iou > self.iou_thresh:
                                # x, y, w, h, scale
                                target[idx, 0, gj, gi] = gx * grid_w - gi
                                target[idx, 1, gj, gi] = gy * grid_h - gj
                                target[idx, 2, gj, gi] = np.log(
                                    gw * w / self.anchors[mask_i][0])
                                target[idx, 3, gj, gi] = np.log(
                                    gh * h / self.anchors[mask_i][1])
                                target[idx, 4, gj, gi] = 2.0 - gw * gh

                                # objectness record gt_score
                                target[idx, 5, gj, gi] = score

                                # classification
                                target[idx, 6 + cls, gj, gi] = 1.
K
Kaipeng Deng 已提交
295 296
                sample['target{}'.format(i)] = target
        return samples
297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457


@register_op
class Gt2FCOSTarget(BaseOperator):
    """
    Generate FCOS targets by groud truth data
    """

    def __init__(self,
                 object_sizes_boundary,
                 center_sampling_radius,
                 downsample_ratios,
                 norm_reg_targets=False):
        super(Gt2FCOSTarget, self).__init__()
        self.center_sampling_radius = center_sampling_radius
        self.downsample_ratios = downsample_ratios
        self.INF = np.inf
        self.object_sizes_boundary = [-1] + object_sizes_boundary + [self.INF]
        object_sizes_of_interest = []
        for i in range(len(self.object_sizes_boundary) - 1):
            object_sizes_of_interest.append([
                self.object_sizes_boundary[i], self.object_sizes_boundary[i + 1]
            ])
        self.object_sizes_of_interest = object_sizes_of_interest
        self.norm_reg_targets = norm_reg_targets

    def _compute_points(self, w, h):
        """
        compute the corresponding points in each feature map
        :param h: image height
        :param w: image width
        :return: points from all feature map
        """
        locations = []
        for stride in self.downsample_ratios:
            shift_x = np.arange(0, w, stride).astype(np.float32)
            shift_y = np.arange(0, h, stride).astype(np.float32)
            shift_x, shift_y = np.meshgrid(shift_x, shift_y)
            shift_x = shift_x.flatten()
            shift_y = shift_y.flatten()
            location = np.stack([shift_x, shift_y], axis=1) + stride // 2
            locations.append(location)
        num_points_each_level = [len(location) for location in locations]
        locations = np.concatenate(locations, axis=0)
        return locations, num_points_each_level

    def _convert_xywh2xyxy(self, gt_bbox, w, h):
        """
        convert the bounding box from style xywh to xyxy
        :param gt_bbox: bounding boxes normalized into [0, 1]
        :param w: image width
        :param h: image height
        :return: bounding boxes in xyxy style
        """
        bboxes = gt_bbox.copy()
        bboxes[:, [0, 2]] = bboxes[:, [0, 2]] * w
        bboxes[:, [1, 3]] = bboxes[:, [1, 3]] * h
        bboxes[:, 2] = bboxes[:, 0] + bboxes[:, 2]
        bboxes[:, 3] = bboxes[:, 1] + bboxes[:, 3]
        return bboxes

    def _check_inside_boxes_limited(self, gt_bbox, xs, ys,
                                    num_points_each_level):
        """
        check if points is within the clipped boxes
        :param gt_bbox: bounding boxes
        :param xs: horizontal coordinate of points
        :param ys: vertical coordinate of points
        :return: the mask of points is within gt_box or not
        """
        bboxes = np.reshape(
            gt_bbox, newshape=[1, gt_bbox.shape[0], gt_bbox.shape[1]])
        bboxes = np.tile(bboxes, reps=[xs.shape[0], 1, 1])
        ct_x = (bboxes[:, :, 0] + bboxes[:, :, 2]) / 2
        ct_y = (bboxes[:, :, 1] + bboxes[:, :, 3]) / 2
        beg = 0
        clipped_box = bboxes.copy()
        for lvl, stride in enumerate(self.downsample_ratios):
            end = beg + num_points_each_level[lvl]
            stride_exp = self.center_sampling_radius * stride
            clipped_box[beg:end, :, 0] = np.maximum(
                bboxes[beg:end, :, 0], ct_x[beg:end, :] - stride_exp)
            clipped_box[beg:end, :, 1] = np.maximum(
                bboxes[beg:end, :, 1], ct_y[beg:end, :] - stride_exp)
            clipped_box[beg:end, :, 2] = np.minimum(
                bboxes[beg:end, :, 2], ct_x[beg:end, :] + stride_exp)
            clipped_box[beg:end, :, 3] = np.minimum(
                bboxes[beg:end, :, 3], ct_y[beg:end, :] + stride_exp)
            beg = end
        l_res = xs - clipped_box[:, :, 0]
        r_res = clipped_box[:, :, 2] - xs
        t_res = ys - clipped_box[:, :, 1]
        b_res = clipped_box[:, :, 3] - ys
        clipped_box_reg_targets = np.stack([l_res, t_res, r_res, b_res], axis=2)
        inside_gt_box = np.min(clipped_box_reg_targets, axis=2) > 0
        return inside_gt_box

    def __call__(self, samples, context=None):
        assert len(self.object_sizes_of_interest) == len(self.downsample_ratios), \
            "object_sizes_of_interest', and 'downsample_ratios' should have same length."

        for sample in samples:
            # im, gt_bbox, gt_class, gt_score = sample
            im = sample['image']
            im_info = sample['im_info']
            bboxes = sample['gt_bbox']
            gt_class = sample['gt_class']
            gt_score = sample['gt_score']
            bboxes[:, [0, 2]] = bboxes[:, [0, 2]] * np.floor(im_info[1]) / \
                np.floor(im_info[1] / im_info[2])
            bboxes[:, [1, 3]] = bboxes[:, [1, 3]] * np.floor(im_info[0]) / \
                np.floor(im_info[0] / im_info[2])
            # calculate the locations
            h, w = sample['image'].shape[1:3]
            points, num_points_each_level = self._compute_points(w, h)
            object_scale_exp = []
            for i, num_pts in enumerate(num_points_each_level):
                object_scale_exp.append(
                    np.tile(
                        np.array([self.object_sizes_of_interest[i]]),
                        reps=[num_pts, 1]))
            object_scale_exp = np.concatenate(object_scale_exp, axis=0)

            gt_area = (bboxes[:, 2] - bboxes[:, 0]) * (
                bboxes[:, 3] - bboxes[:, 1])
            xs, ys = points[:, 0], points[:, 1]
            xs = np.reshape(xs, newshape=[xs.shape[0], 1])
            xs = np.tile(xs, reps=[1, bboxes.shape[0]])
            ys = np.reshape(ys, newshape=[ys.shape[0], 1])
            ys = np.tile(ys, reps=[1, bboxes.shape[0]])

            l_res = xs - bboxes[:, 0]
            r_res = bboxes[:, 2] - xs
            t_res = ys - bboxes[:, 1]
            b_res = bboxes[:, 3] - ys
            reg_targets = np.stack([l_res, t_res, r_res, b_res], axis=2)
            if self.center_sampling_radius > 0:
                is_inside_box = self._check_inside_boxes_limited(
                    bboxes, xs, ys, num_points_each_level)
            else:
                is_inside_box = np.min(reg_targets, axis=2) > 0
            # check if the targets is inside the corresponding level
            max_reg_targets = np.max(reg_targets, axis=2)
            lower_bound = np.tile(
                np.expand_dims(
                    object_scale_exp[:, 0], axis=1),
                reps=[1, max_reg_targets.shape[1]])
            high_bound = np.tile(
                np.expand_dims(
                    object_scale_exp[:, 1], axis=1),
                reps=[1, max_reg_targets.shape[1]])
            is_match_current_level = \
                (max_reg_targets > lower_bound) & \
                (max_reg_targets < high_bound)
            points2gtarea = np.tile(
                np.expand_dims(
                    gt_area, axis=0), reps=[xs.shape[0], 1])
            points2gtarea[is_inside_box == 0] = self.INF
            points2gtarea[is_match_current_level == 0] = self.INF
            points2min_area = points2gtarea.min(axis=1)
            points2min_area_ind = points2gtarea.argmin(axis=1)
458
            labels = gt_class[points2min_area_ind] + 1
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
            labels[points2min_area == self.INF] = 0
            reg_targets = reg_targets[range(xs.shape[0]), points2min_area_ind]
            ctn_targets = np.sqrt((reg_targets[:, [0, 2]].min(axis=1) / \
                                  reg_targets[:, [0, 2]].max(axis=1)) * \
                                  (reg_targets[:, [1, 3]].min(axis=1) / \
                                   reg_targets[:, [1, 3]].max(axis=1))).astype(np.float32)
            ctn_targets = np.reshape(
                ctn_targets, newshape=[ctn_targets.shape[0], 1])
            ctn_targets[labels <= 0] = 0
            pos_ind = np.nonzero(labels != 0)
            reg_targets_pos = reg_targets[pos_ind[0], :]
            split_sections = []
            beg = 0
            for lvl in range(len(num_points_each_level)):
                end = beg + num_points_each_level[lvl]
                split_sections.append(end)
                beg = end
            labels_by_level = np.split(labels, split_sections, axis=0)
            reg_targets_by_level = np.split(reg_targets, split_sections, axis=0)
            ctn_targets_by_level = np.split(ctn_targets, split_sections, axis=0)
            for lvl in range(len(self.downsample_ratios)):
                grid_w = int(np.ceil(w / self.downsample_ratios[lvl]))
                grid_h = int(np.ceil(h / self.downsample_ratios[lvl]))
                if self.norm_reg_targets:
                    sample['reg_target{}'.format(lvl)] = \
                        np.reshape(
                            reg_targets_by_level[lvl] / \
                            self.downsample_ratios[lvl],
                            newshape=[grid_h, grid_w, 4])
                else:
                    sample['reg_target{}'.format(lvl)] = np.reshape(
                        reg_targets_by_level[lvl],
                        newshape=[grid_h, grid_w, 4])
                sample['labels{}'.format(lvl)] = np.reshape(
                    labels_by_level[lvl], newshape=[grid_h, grid_w, 1])
                sample['centerness{}'.format(lvl)] = np.reshape(
                    ctn_targets_by_level[lvl], newshape=[grid_h, grid_w, 1])
        return samples
W
wangguanzhong 已提交
497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592


@register_op
class Gt2TTFTarget(BaseOperator):
    """
    Gt2TTFTarget
    Generate TTFNet targets by ground truth data
    
    Args:
        num_classes(int): the number of classes.
        down_ratio(int): the down ratio from images to heatmap, 4 by default.
        alpha(float): the alpha parameter to generate gaussian target.
            0.54 by default.
    """

    def __init__(self, num_classes, down_ratio=4, alpha=0.54):
        super(Gt2TTFTarget, self).__init__()
        self.down_ratio = down_ratio
        self.num_classes = num_classes
        self.alpha = alpha

    def __call__(self, samples, context=None):
        output_size = samples[0]['image'].shape[1]
        feat_size = output_size // self.down_ratio
        for sample in samples:
            heatmap = np.zeros(
                (self.num_classes, feat_size, feat_size), dtype='float32')
            box_target = np.ones(
                (4, feat_size, feat_size), dtype='float32') * -1
            reg_weight = np.zeros((1, feat_size, feat_size), dtype='float32')

            gt_bbox = sample['gt_bbox']
            gt_class = sample['gt_class']

            bbox_w = gt_bbox[:, 2] - gt_bbox[:, 0] + 1
            bbox_h = gt_bbox[:, 3] - gt_bbox[:, 1] + 1
            area = bbox_w * bbox_h
            boxes_areas_log = np.log(area)
            boxes_ind = np.argsort(boxes_areas_log, axis=0)[::-1]
            boxes_area_topk_log = boxes_areas_log[boxes_ind]
            gt_bbox = gt_bbox[boxes_ind]
            gt_class = gt_class[boxes_ind]

            feat_gt_bbox = gt_bbox / self.down_ratio
            feat_gt_bbox = np.clip(feat_gt_bbox, 0, feat_size - 1)
            feat_hs, feat_ws = (feat_gt_bbox[:, 3] - feat_gt_bbox[:, 1],
                                feat_gt_bbox[:, 2] - feat_gt_bbox[:, 0])

            ct_inds = np.stack(
                [(gt_bbox[:, 0] + gt_bbox[:, 2]) / 2,
                 (gt_bbox[:, 1] + gt_bbox[:, 3]) / 2],
                axis=1) / self.down_ratio

            h_radiuses_alpha = (feat_hs / 2. * self.alpha).astype('int32')
            w_radiuses_alpha = (feat_ws / 2. * self.alpha).astype('int32')

            for k in range(len(gt_bbox)):
                cls_id = gt_class[k]
                fake_heatmap = np.zeros((feat_size, feat_size), dtype='float32')
                self.draw_truncate_gaussian(fake_heatmap, ct_inds[k],
                                            h_radiuses_alpha[k],
                                            w_radiuses_alpha[k])

                heatmap[cls_id] = np.maximum(heatmap[cls_id], fake_heatmap)
                box_target_inds = fake_heatmap > 0
                box_target[:, box_target_inds] = gt_bbox[k][:, None]

                local_heatmap = fake_heatmap[box_target_inds]
                ct_div = np.sum(local_heatmap)
                local_heatmap *= boxes_area_topk_log[k]
                reg_weight[0, box_target_inds] = local_heatmap / ct_div
            sample['ttf_heatmap'] = heatmap
            sample['ttf_box_target'] = box_target
            sample['ttf_reg_weight'] = reg_weight
        return samples

    def draw_truncate_gaussian(self, heatmap, center, h_radius, w_radius):
        h, w = 2 * h_radius + 1, 2 * w_radius + 1
        sigma_x = w / 6
        sigma_y = h / 6
        gaussian = gaussian2D((h, w), sigma_x, sigma_y)

        x, y = int(center[0]), int(center[1])

        height, width = heatmap.shape[0:2]

        left, right = min(x, w_radius), min(width - x, w_radius + 1)
        top, bottom = min(y, h_radius), min(height - y, h_radius + 1)

        masked_heatmap = heatmap[y - top:y + bottom, x - left:x + right]
        masked_gaussian = gaussian[h_radius - top:h_radius + bottom, w_radius -
                                   left:w_radius + right]
        if min(masked_gaussian.shape) > 0 and min(masked_heatmap.shape) > 0:
            heatmap[y - top:y + bottom, x - left:x + right] = np.maximum(
                masked_heatmap, masked_gaussian)
        return heatmap