op_helper.py 17.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
# 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.
# this file contains helper methods for BBOX processing

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

import numpy as np
21 22 23
import random
import math
import cv2
24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45


def meet_emit_constraint(src_bbox, sample_bbox):
    center_x = (src_bbox[2] + src_bbox[0]) / 2
    center_y = (src_bbox[3] + src_bbox[1]) / 2
    if center_x >= sample_bbox[0] and \
            center_x <= sample_bbox[2] and \
            center_y >= sample_bbox[1] and \
            center_y <= sample_bbox[3]:
        return True
    return False


def clip_bbox(src_bbox):
    src_bbox[0] = max(min(src_bbox[0], 1.0), 0.0)
    src_bbox[1] = max(min(src_bbox[1], 1.0), 0.0)
    src_bbox[2] = max(min(src_bbox[2], 1.0), 0.0)
    src_bbox[3] = max(min(src_bbox[3], 1.0), 0.0)
    return src_bbox


def bbox_area(src_bbox):
46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61
    if src_bbox[2] < src_bbox[0] or src_bbox[3] < src_bbox[1]:
        return 0.
    else:
        width = src_bbox[2] - src_bbox[0]
        height = src_bbox[3] - src_bbox[1]
        return width * height


def is_overlap(object_bbox, sample_bbox):
    if object_bbox[0] >= sample_bbox[2] or \
       object_bbox[2] <= sample_bbox[0] or \
       object_bbox[1] >= sample_bbox[3] or \
       object_bbox[3] <= sample_bbox[1]:
        return False
    else:
        return True
62 63


W
wangxinxin08 已提交
64 65 66 67
def filter_and_process(sample_bbox,
                       bboxes,
                       labels,
                       scores=None,
68
                       keypoints=None):
69 70 71
    new_bboxes = []
    new_labels = []
    new_scores = []
72 73
    new_keypoints = []
    new_kp_ignore = []
74
    for i in range(len(bboxes)):
75 76 77 78
        new_bbox = [0, 0, 0, 0]
        obj_bbox = [bboxes[i][0], bboxes[i][1], bboxes[i][2], bboxes[i][3]]
        if not meet_emit_constraint(obj_bbox, sample_bbox):
            continue
79 80
        if not is_overlap(obj_bbox, sample_bbox):
            continue
81 82 83 84 85 86 87 88 89 90 91 92
        sample_width = sample_bbox[2] - sample_bbox[0]
        sample_height = sample_bbox[3] - sample_bbox[1]
        new_bbox[0] = (obj_bbox[0] - sample_bbox[0]) / sample_width
        new_bbox[1] = (obj_bbox[1] - sample_bbox[1]) / sample_height
        new_bbox[2] = (obj_bbox[2] - sample_bbox[0]) / sample_width
        new_bbox[3] = (obj_bbox[3] - sample_bbox[1]) / sample_height
        new_bbox = clip_bbox(new_bbox)
        if bbox_area(new_bbox) > 0:
            new_bboxes.append(new_bbox)
            new_labels.append([labels[i][0]])
            if scores is not None:
                new_scores.append([scores[i][0]])
93 94 95 96 97
            if keypoints is not None:
                sample_keypoint = keypoints[0][i]
                for j in range(len(sample_keypoint)):
                    kp_len = sample_height if j % 2 else sample_width
                    sample_coord = sample_bbox[1] if j % 2 else sample_bbox[0]
W
wangxinxin08 已提交
98 99
                    sample_keypoint[j] = (sample_keypoint[j] -
                                          sample_coord) / kp_len
100 101 102 103
                    sample_keypoint[j] = max(min(sample_keypoint[j], 1.0), 0.0)
                new_keypoints.append(sample_keypoint)
                new_kp_ignore.append(keypoints[1][i])

104 105 106
    bboxes = np.array(new_bboxes)
    labels = np.array(new_labels)
    scores = np.array(new_scores)
107 108 109 110
    if keypoints is not None:
        keypoints = np.array(new_keypoints)
        new_kp_ignore = np.array(new_kp_ignore)
        return bboxes, labels, scores, (keypoints, new_kp_ignore)
111 112 113
    return bboxes, labels, scores


114 115 116 117 118 119 120 121 122 123 124 125
def bbox_area_sampling(bboxes, labels, scores, target_size, min_size):
    new_bboxes = []
    new_labels = []
    new_scores = []
    for i, bbox in enumerate(bboxes):
        w = float((bbox[2] - bbox[0]) * target_size)
        h = float((bbox[3] - bbox[1]) * target_size)
        if w * h < float(min_size * min_size):
            continue
        else:
            new_bboxes.append(bbox)
            new_labels.append(labels[i])
126 127
            if scores is not None and scores.size != 0:
                new_scores.append(scores[i])
128 129 130 131 132 133
    bboxes = np.array(new_bboxes)
    labels = np.array(new_labels)
    scores = np.array(new_scores)
    return bboxes, labels, scores


134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150
def generate_sample_bbox(sampler):
    scale = np.random.uniform(sampler[2], sampler[3])
    aspect_ratio = np.random.uniform(sampler[4], sampler[5])
    aspect_ratio = max(aspect_ratio, (scale**2.0))
    aspect_ratio = min(aspect_ratio, 1 / (scale**2.0))
    bbox_width = scale * (aspect_ratio**0.5)
    bbox_height = scale / (aspect_ratio**0.5)
    xmin_bound = 1 - bbox_width
    ymin_bound = 1 - bbox_height
    xmin = np.random.uniform(0, xmin_bound)
    ymin = np.random.uniform(0, ymin_bound)
    xmax = xmin + bbox_width
    ymax = ymin + bbox_height
    sampled_bbox = [xmin, ymin, xmax, ymax]
    return sampled_bbox


151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256
def generate_sample_bbox_square(sampler, image_width, image_height):
    scale = np.random.uniform(sampler[2], sampler[3])
    aspect_ratio = np.random.uniform(sampler[4], sampler[5])
    aspect_ratio = max(aspect_ratio, (scale**2.0))
    aspect_ratio = min(aspect_ratio, 1 / (scale**2.0))
    bbox_width = scale * (aspect_ratio**0.5)
    bbox_height = scale / (aspect_ratio**0.5)
    if image_height < image_width:
        bbox_width = bbox_height * image_height / image_width
    else:
        bbox_height = bbox_width * image_width / image_height
    xmin_bound = 1 - bbox_width
    ymin_bound = 1 - bbox_height
    xmin = np.random.uniform(0, xmin_bound)
    ymin = np.random.uniform(0, ymin_bound)
    xmax = xmin + bbox_width
    ymax = ymin + bbox_height
    sampled_bbox = [xmin, ymin, xmax, ymax]
    return sampled_bbox


def data_anchor_sampling(bbox_labels, image_width, image_height, scale_array,
                         resize_width):
    num_gt = len(bbox_labels)
    # np.random.randint range: [low, high)
    rand_idx = np.random.randint(0, num_gt) if num_gt != 0 else 0

    if num_gt != 0:
        norm_xmin = bbox_labels[rand_idx][0]
        norm_ymin = bbox_labels[rand_idx][1]
        norm_xmax = bbox_labels[rand_idx][2]
        norm_ymax = bbox_labels[rand_idx][3]

        xmin = norm_xmin * image_width
        ymin = norm_ymin * image_height
        wid = image_width * (norm_xmax - norm_xmin)
        hei = image_height * (norm_ymax - norm_ymin)
        range_size = 0

        area = wid * hei
        for scale_ind in range(0, len(scale_array) - 1):
            if area > scale_array[scale_ind] ** 2 and area < \
                    scale_array[scale_ind + 1] ** 2:
                range_size = scale_ind + 1
                break

        if area > scale_array[len(scale_array) - 2]**2:
            range_size = len(scale_array) - 2

        scale_choose = 0.0
        if range_size == 0:
            rand_idx_size = 0
        else:
            # np.random.randint range: [low, high)
            rng_rand_size = np.random.randint(0, range_size + 1)
            rand_idx_size = rng_rand_size % (range_size + 1)

        if rand_idx_size == range_size:
            min_resize_val = scale_array[rand_idx_size] / 2.0
            max_resize_val = min(2.0 * scale_array[rand_idx_size],
                                 2 * math.sqrt(wid * hei))
            scale_choose = random.uniform(min_resize_val, max_resize_val)
        else:
            min_resize_val = scale_array[rand_idx_size] / 2.0
            max_resize_val = 2.0 * scale_array[rand_idx_size]
            scale_choose = random.uniform(min_resize_val, max_resize_val)

        sample_bbox_size = wid * resize_width / scale_choose

        w_off_orig = 0.0
        h_off_orig = 0.0
        if sample_bbox_size < max(image_height, image_width):
            if wid <= sample_bbox_size:
                w_off_orig = np.random.uniform(xmin + wid - sample_bbox_size,
                                               xmin)
            else:
                w_off_orig = np.random.uniform(xmin,
                                               xmin + wid - sample_bbox_size)

            if hei <= sample_bbox_size:
                h_off_orig = np.random.uniform(ymin + hei - sample_bbox_size,
                                               ymin)
            else:
                h_off_orig = np.random.uniform(ymin,
                                               ymin + hei - sample_bbox_size)

        else:
            w_off_orig = np.random.uniform(image_width - sample_bbox_size, 0.0)
            h_off_orig = np.random.uniform(image_height - sample_bbox_size, 0.0)

        w_off_orig = math.floor(w_off_orig)
        h_off_orig = math.floor(h_off_orig)

        # Figure out top left coordinates.
        w_off = float(w_off_orig / image_width)
        h_off = float(h_off_orig / image_height)

        sampled_bbox = [
            w_off, h_off, w_off + float(sample_bbox_size / image_width),
            h_off + float(sample_bbox_size / image_height)
        ]
        return sampled_bbox
    else:
        return 0


257 258 259 260 261 262 263 264 265 266
def jaccard_overlap(sample_bbox, object_bbox):
    if sample_bbox[0] >= object_bbox[2] or \
        sample_bbox[2] <= object_bbox[0] or \
        sample_bbox[1] >= object_bbox[3] or \
        sample_bbox[3] <= object_bbox[1]:
        return 0
    intersect_xmin = max(sample_bbox[0], object_bbox[0])
    intersect_ymin = max(sample_bbox[1], object_bbox[1])
    intersect_xmax = min(sample_bbox[2], object_bbox[2])
    intersect_ymax = min(sample_bbox[3], object_bbox[3])
W
wangxinxin08 已提交
267 268
    intersect_size = (intersect_xmax - intersect_xmin) * (intersect_ymax -
                                                          intersect_ymin)
269 270
    sample_bbox_size = bbox_area(sample_bbox)
    object_bbox_size = bbox_area(object_bbox)
W
wangxinxin08 已提交
271 272
    overlap = intersect_size / (sample_bbox_size + object_bbox_size -
                                intersect_size)
273 274 275
    return overlap


276 277 278 279 280 281
def intersect_bbox(bbox1, bbox2):
    if bbox2[0] > bbox1[2] or bbox2[2] < bbox1[0] or \
        bbox2[1] > bbox1[3] or bbox2[3] < bbox1[1]:
        intersection_box = [0.0, 0.0, 0.0, 0.0]
    else:
        intersection_box = [
W
wangxinxin08 已提交
282 283 284 285
            max(bbox1[0], bbox2[0]),
            max(bbox1[1], bbox2[1]),
            min(bbox1[2], bbox2[2]),
            min(bbox1[3], bbox2[3])
286 287 288 289 290 291 292 293 294 295 296 297 298 299 300
        ]
    return intersection_box


def bbox_coverage(bbox1, bbox2):
    inter_box = intersect_bbox(bbox1, bbox2)
    intersect_size = bbox_area(inter_box)

    if intersect_size > 0:
        bbox1_size = bbox_area(bbox1)
        return intersect_size / bbox1_size
    else:
        return 0.


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
def satisfy_sample_constraint(sampler,
                              sample_bbox,
                              gt_bboxes,
                              satisfy_all=False):
    if sampler[6] == 0 and sampler[7] == 0:
        return True
    satisfied = []
    for i in range(len(gt_bboxes)):
        object_bbox = [
            gt_bboxes[i][0], gt_bboxes[i][1], gt_bboxes[i][2], gt_bboxes[i][3]
        ]
        overlap = jaccard_overlap(sample_bbox, object_bbox)
        if sampler[6] != 0 and \
                overlap < sampler[6]:
            satisfied.append(False)
            continue
        if sampler[7] != 0 and \
                overlap > sampler[7]:
            satisfied.append(False)
            continue
        satisfied.append(True)
        if not satisfy_all:
            return True

    if satisfy_all:
        return np.all(satisfied)
    else:
        return False
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


def satisfy_sample_constraint_coverage(sampler, sample_bbox, gt_bboxes):
    if sampler[6] == 0 and sampler[7] == 0:
        has_jaccard_overlap = False
    else:
        has_jaccard_overlap = True
    if sampler[8] == 0 and sampler[9] == 0:
        has_object_coverage = False
    else:
        has_object_coverage = True

    if not has_jaccard_overlap and not has_object_coverage:
        return True
    found = False
    for i in range(len(gt_bboxes)):
        object_bbox = [
            gt_bboxes[i][0], gt_bboxes[i][1], gt_bboxes[i][2], gt_bboxes[i][3]
        ]
        if has_jaccard_overlap:
            overlap = jaccard_overlap(sample_bbox, object_bbox)
            if sampler[6] != 0 and \
                    overlap < sampler[6]:
                continue
            if sampler[7] != 0 and \
                    overlap > sampler[7]:
                continue
            found = True
        if has_object_coverage:
            object_coverage = bbox_coverage(object_bbox, sample_bbox)
            if sampler[8] != 0 and \
                    object_coverage < sampler[8]:
                continue
            if sampler[9] != 0 and \
                    object_coverage > sampler[9]:
                continue
            found = True
        if found:
            return True
    return found


def crop_image_sampling(img, sample_bbox, image_width, image_height,
                        target_size):
    # no clipping here
    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)

    w_off = xmin
    h_off = ymin
    width = xmax - xmin
    height = ymax - ymin
    cross_xmin = max(0.0, float(w_off))
    cross_ymin = max(0.0, float(h_off))
    cross_xmax = min(float(w_off + width - 1.0), float(image_width))
    cross_ymax = min(float(h_off + height - 1.0), float(image_height))
    cross_width = cross_xmax - cross_xmin
    cross_height = cross_ymax - cross_ymin

    roi_xmin = 0 if w_off >= 0 else abs(w_off)
    roi_ymin = 0 if h_off >= 0 else abs(h_off)
    roi_width = cross_width
    roi_height = cross_height

    roi_y1 = int(roi_ymin)
    roi_y2 = int(roi_ymin + roi_height)
    roi_x1 = int(roi_xmin)
    roi_x2 = int(roi_xmin + roi_width)

    cross_y1 = int(cross_ymin)
    cross_y2 = int(cross_ymin + cross_height)
    cross_x1 = int(cross_xmin)
    cross_x2 = int(cross_xmin + cross_width)

    sample_img = np.zeros((height, width, 3))
    sample_img[roi_y1: roi_y2, roi_x1: roi_x2] = \
        img[cross_y1: cross_y2, cross_x1: cross_x2]

W
wangxinxin08 已提交
409 410
    sample_img = cv2.resize(sample_img, (target_size, target_size),
                            interpolation=cv2.INTER_AREA)
411 412

    return sample_img
413 414 415 416 417 418


def is_poly(segm):
    assert isinstance(segm, (list, dict)), \
        "Invalid segm type: {}".format(type(segm))
    return isinstance(segm, list)
W
wangguanzhong 已提交
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


def gaussian_radius(bbox_size, min_overlap):
    height, width = bbox_size

    a1 = 1
    b1 = (height + width)
    c1 = width * height * (1 - min_overlap) / (1 + min_overlap)
    sq1 = np.sqrt(b1**2 - 4 * a1 * c1)
    radius1 = (b1 - sq1) / (2 * a1)

    a2 = 4
    b2 = 2 * (height + width)
    c2 = (1 - min_overlap) * width * height
    sq2 = np.sqrt(b2**2 - 4 * a2 * c2)
    radius2 = (b2 - sq2) / (2 * a2)

    a3 = 4 * min_overlap
    b3 = -2 * min_overlap * (height + width)
    c3 = (min_overlap - 1) * width * height
    sq3 = np.sqrt(b3**2 - 4 * a3 * c3)
    radius3 = (b3 + sq3) / (2 * a3)
    return min(radius1, radius2, radius3)


def draw_gaussian(heatmap, center, radius, k=1, delte=6):
    diameter = 2 * radius + 1
W
wangguanzhong 已提交
446 447
    sigma = diameter / delte
    gaussian = gaussian2D((diameter, diameter), sigma_x=sigma, sigma_y=sigma)
W
wangguanzhong 已提交
448 449 450 451 452 453 454 455 456

    x, y = center

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

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

    masked_heatmap = heatmap[y - top:y + bottom, x - left:x + right]
W
wangxinxin08 已提交
457 458
    masked_gaussian = gaussian[radius - top:radius + bottom,
                               radius - left:radius + right]
W
wangguanzhong 已提交
459 460 461
    np.maximum(masked_heatmap, masked_gaussian * k, out=masked_heatmap)


W
wangguanzhong 已提交
462
def gaussian2D(shape, sigma_x=1, sigma_y=1):
W
wangguanzhong 已提交
463 464 465
    m, n = [(ss - 1.) / 2. for ss in shape]
    y, x = np.ogrid[-m:m + 1, -n:n + 1]

W
wangxinxin08 已提交
466 467
    h = np.exp(-(x * x / (2 * sigma_x * sigma_x) + y * y /
                 (2 * sigma_y * sigma_y)))
W
wangguanzhong 已提交
468 469
    h[h < np.finfo(h.dtype).eps * h.max()] = 0
    return h
W
wangxinxin08 已提交
470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515


def transform_bbox(bbox,
                   label,
                   M,
                   w,
                   h,
                   area_thr=0.25,
                   wh_thr=2,
                   ar_thr=20,
                   perspective=False):
    # rotate bbox
    n = len(bbox)
    xy = np.ones((n * 4, 3), dtype=np.float32)
    xy[:, :2] = bbox[:, [0, 1, 2, 3, 0, 3, 2, 1]].reshape(n * 4, 2)
    xy = xy @ M.T
    if perspective:
        xy = (xy[:, :2] / xy[:, 2:3]).reshape(n, 8)
    else:
        xy = xy[:, :2].reshape(n, 8)
    # get new bboxes
    x = xy[:, [0, 2, 4, 6]]
    y = xy[:, [1, 3, 5, 7]]
    new_bbox = np.concatenate(
        (x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T
    # clip boxes
    new_bbox, mask = clip_bbox(new_bbox, w, h, area_thr)
    new_label = label[mask]
    return new_bbox, new_label


def clip_bbox(bbox, w, h, area_thr=0.25, wh_thr=2, ar_thr=20):
    # clip boxes
    area1 = (bbox[:, 2:4] - bbox[:, 0:2]).prod(1)
    bbox[:, [0, 2]] = bbox[:, [0, 2]].clip(0, w)
    bbox[:, [1, 3]] = bbox[:, [1, 3]].clip(0, h)
    # compute
    area2 = (bbox[:, 2:4] - bbox[:, 0:2]).prod(1)
    area_ratio = area2 / (area1 + 1e-16)
    wh = bbox[:, 2:4] - bbox[:, 0:2]
    ar_ratio = np.maximum(wh[:, 1] / (wh[:, 0] + 1e-16),
                          wh[:, 0] / (wh[:, 1] + 1e-16))
    mask = (area_ratio > area_thr) & (
        (wh > wh_thr).all(1)) & (ar_ratio < ar_thr)
    bbox = bbox[mask]
    return bbox, mask