data_utils.py 15 KB
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 29 30 31 32 33 34 35 36 37 38 39 40 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 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430
from __future__ import division, print_function

import tensorflow as tf
import numpy as np
import cv2
import sys
import random

PY_VERSION = sys.version_info[0]
iter_cnt = 0

FEATURE_DESCRIPTION = {
    'index': tf.FixedLenFeature([], tf.int64),
    'image': tf.FixedLenFeature([], tf.string),
    'width': tf.FixedLenFeature([], tf.int64),
    'height': tf.FixedLenFeature([], tf.int64),
    'boxes': tf.VarLenFeature(tf.int64)
}

def parse_tfrecord(data):
    # tfrecord parser for TFRecordDataset (raw data)
    features = tf.parse_single_example(data, FEATURE_DESCRIPTION)
    index = int(features['index'])
    encoded_image = np.frombuffer(features['image'], dtype = np.uint8)
    width = int(features['width'])
    height = int(features['height'])
    boxes = features['boxes'].eval()

    assert len(boxes) % 5 == 0, 'Annotation error occured in box array.'
    box_cnt = len(boxes) // 5

    aligned_boxes = []
    labels = []

    for i in range(box_cnt):
        label, x_min, y_min, x_max, y_max = int(boxes[i * 5]), float(boxes[i * 5 + 1]), float(boxes[i * 5 + 2]), float(boxes[i * 5 + 3]) ## do we need to change int to float? is there float rectangle sample?
        aligned_boxes.append([x_min, y_min, x_max, y_max])
        labels.append(label)

    aligned_boxes = np.asarray(aligned_boxes, np.float32)
    labels = np.asarray(labels, np.int64)

    return index, encoded_image, aligned_boxes, labels, width, height

def parse_record(features):
    # tfrecord parser for TFRecordIterator (primitive data)

    index = int(features['index'])
    encoded_image = np.frombuffer(features['image'], dtype = np.uint8)
    width = int(features['width'])
    height = int(features['height'])
    boxes = features['boxes']

    assert len(boxes) % 5 == 0, 'Annotation error occured in box array.'
    box_cnt = len(boxes) // 5

    aligned_boxes = []
    labels = []

    for i in range(box_cnt):
        label, x_min, y_min, x_max, y_max = int(boxes[i * 5]), float(boxes[i * 5 + 1]), float(boxes[i * 5 + 2]), float(boxes[i * 5 + 3])
        aligned_boxes.append([x_min, y_min, x_max, y_max])
        labels.append(label)

    aligned_boxes = np.asarray(aligned_boxes, np.float32)
    labels = np.asarray(labels, np.int64)

    return index, encoded_image, aligned_boxes, labels, width, height

def bbox_crop(bbox, crop_box=None, allow_outside_center=True):
    bbox = bbox.copy()
    if crop_box is None:
        return bbox
    if not len(crop_box) == 4:
        raise ValueError(
            "Invalid crop_box parameter, requires length 4, given {}".format(str(crop_box)))
    if sum([int(c is None) for c in crop_box]) == 4:
        return bbox

    l, t, w, h = crop_box

    left = l if l else 0
    top = t if t else 0
    right = left + (w if w else np.inf)
    bottom = top + (h if h else np.inf)
    crop_bbox = np.array((left, top, right, bottom))

    if allow_outside_center:
        mask = np.ones(bbox.shape[0], dtype=bool)
    else:
        centers = (bbox[:, :2] + bbox[:, 2:4]) / 2
        mask = np.logical_and(crop_bbox[:2] <= centers, centers < crop_bbox[2:]).all(axis=1)

    # transform borders
    bbox[:, :2] = np.maximum(bbox[:, :2], crop_bbox[:2])
    bbox[:, 2:4] = np.minimum(bbox[:, 2:4], crop_bbox[2:4])
    bbox[:, :2] -= crop_bbox[:2]
    bbox[:, 2:4] -= crop_bbox[:2]

    mask = np.logical_and(mask, (bbox[:, :2] < bbox[:, 2:4]).all(axis=1))
    bbox = bbox[mask]
    return bbox

def bbox_iou(bbox_a, bbox_b, offset=0):
    if bbox_a.shape[1] < 4 or bbox_b.shape[1] < 4:
        raise IndexError("Bounding boxes axis 1 must have at least length 4")

    tl = np.maximum(bbox_a[:, None, :2], bbox_b[:, :2])
    br = np.minimum(bbox_a[:, None, 2:4], bbox_b[:, 2:4])

    area_i = np.prod(br - tl + offset, axis=2) * (tl < br).all(axis=2)
    area_a = np.prod(bbox_a[:, 2:4] - bbox_a[:, :2] + offset, axis=1)
    area_b = np.prod(bbox_b[:, 2:4] - bbox_b[:, :2] + offset, axis=1)
    return area_i / (area_a[:, None] + area_b - area_i)


def random_crop_with_constraints(bbox, size, min_scale=0.3, max_scale=1,
                                 max_aspect_ratio=2, constraints=None,
                                 max_trial=50):
    # default params in paper
    if constraints is None:
        constraints = (
            (0.1, None),
            (0.3, None),
            (0.5, None),
            (0.7, None),
            (0.9, None),
            (None, 1),
        )

    w, h = size

    candidates = [(0, 0, w, h)]
    for min_iou, max_iou in constraints:
        min_iou = -np.inf if min_iou is None else min_iou
        max_iou = np.inf if max_iou is None else max_iou

        for _ in range(max_trial):
            scale = random.uniform(min_scale, max_scale)
            aspect_ratio = random.uniform(
                max(1 / max_aspect_ratio, scale * scale),
                min(max_aspect_ratio, 1 / (scale * scale)))
            crop_h = int(h * scale / np.sqrt(aspect_ratio))
            crop_w = int(w * scale * np.sqrt(aspect_ratio))

            crop_t = random.randrange(h - crop_h)
            crop_l = random.randrange(w - crop_w)
            crop_bb = np.array((crop_l, crop_t, crop_l + crop_w, crop_t + crop_h))

            if len(bbox) == 0:
                top, bottom = crop_t, crop_t + crop_h
                left, right = crop_l, crop_l + crop_w
                return bbox, (left, top, right-left, bottom-top)

            iou = bbox_iou(bbox, crop_bb[np.newaxis])
            if min_iou <= iou.min() and iou.max() <= max_iou:
                top, bottom = crop_t, crop_t + crop_h
                left, right = crop_l, crop_l + crop_w
                candidates.append((left, top, right-left, bottom-top))
                break

    # random select one
    while candidates:
        crop = candidates.pop(np.random.randint(0, len(candidates)))
        new_bbox = bbox_crop(bbox, crop, allow_outside_center=False)
        if new_bbox.size < 1:
            continue
        new_crop = (crop[0], crop[1], crop[2], crop[3])
        return new_bbox, new_crop
    return bbox, (0, 0, w, h)


def random_color_distort(img, brightness_delta=32, hue_vari=18, sat_vari=0.5, val_vari=0.5):
    def random_hue(img_hsv, hue_vari, p=0.5):
        if np.random.uniform(0, 1) > p:
            hue_delta = np.random.randint(-hue_vari, hue_vari)
            img_hsv[:, :, 0] = (img_hsv[:, :, 0] + hue_delta) % 180
        return img_hsv

    def random_saturation(img_hsv, sat_vari, p=0.5):
        if np.random.uniform(0, 1) > p:
            sat_mult = 1 + np.random.uniform(-sat_vari, sat_vari)
            img_hsv[:, :, 1] *= sat_mult
        return img_hsv

    def random_value(img_hsv, val_vari, p=0.5):
        if np.random.uniform(0, 1) > p:
            val_mult = 1 + np.random.uniform(-val_vari, val_vari)
            img_hsv[:, :, 2] *= val_mult
        return img_hsv

    def random_brightness(img, brightness_delta, p=0.5):
        if np.random.uniform(0, 1) > p:
            img = img.astype(np.float32)
            brightness_delta = int(np.random.uniform(-brightness_delta, brightness_delta))
            img = img + brightness_delta
        return np.clip(img, 0, 255)

    # brightness
    img = random_brightness(img, brightness_delta)
    img = img.astype(np.uint8)

    # color jitter
    img_hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV).astype(np.float32)

    if np.random.randint(0, 2):
        img_hsv = random_value(img_hsv, val_vari)
        img_hsv = random_saturation(img_hsv, sat_vari)
        img_hsv = random_hue(img_hsv, hue_vari)
    else:
        img_hsv = random_saturation(img_hsv, sat_vari)
        img_hsv = random_hue(img_hsv, hue_vari)
        img_hsv = random_value(img_hsv, val_vari)

    img_hsv = np.clip(img_hsv, 0, 255)
    img = cv2.cvtColor(img_hsv.astype(np.uint8), cv2.COLOR_HSV2BGR)

    return img


def letterbox_resize(img, new_width, new_height, interp=0):
    ori_height, ori_width = img.shape[:2]

    resize_ratio = min(new_width / ori_width, new_height / ori_height)

    resize_w = int(resize_ratio * ori_width)
    resize_h = int(resize_ratio * ori_height)

    img = cv2.resize(img, (resize_w, resize_h), interpolation=interp)
    image_padded = np.full((new_height, new_width, 3), 128, np.uint8)

    dw = int((new_width - resize_w) / 2)
    dh = int((new_height - resize_h) / 2)

    image_padded[dh: resize_h + dh, dw: resize_w + dw, :] = img

    return image_padded, resize_ratio, dw, dh


def resize_with_bbox(img, bbox, new_width, new_height, interp=0, letterbox=False):
    if letterbox:
        image_padded, resize_ratio, dw, dh = letterbox_resize(img, new_width, new_height, interp)

        # xmin, xmax
        bbox[:, [0, 2]] = bbox[:, [0, 2]] * resize_ratio + dw
        # ymin, ymax
        bbox[:, [1, 3]] = bbox[:, [1, 3]] * resize_ratio + dh

        return image_padded, bbox
    else:
        ori_height, ori_width = img.shape[:2]

        img = cv2.resize(img, (new_width, new_height), interpolation=interp)

        # xmin, xmax
        bbox[:, [0, 2]] = bbox[:, [0, 2]] / ori_width * new_width
        # ymin, ymax
        bbox[:, [1, 3]] = bbox[:, [1, 3]] / ori_height * new_height

        return img, bbox


def random_flip(img, bbox, px=0, py=0):
    height, width = img.shape[:2]
    if np.random.uniform(0, 1) < px:
        img = cv2.flip(img, 1)
        xmax = width - bbox[:, 0]
        xmin = width - bbox[:, 2]
        bbox[:, 0] = xmin
        bbox[:, 2] = xmax

    if np.random.uniform(0, 1) < py:
        img = cv2.flip(img, 0)
        ymax = height - bbox[:, 1]
        ymin = height - bbox[:, 3]
        bbox[:, 1] = ymin
        bbox[:, 3] = ymax
    return img, bbox


def random_expand(img, bbox, max_ratio=4, fill=0, keep_ratio=True):
    h, w, c = img.shape
    ratio_x = random.uniform(1, max_ratio)
    if keep_ratio:
        ratio_y = ratio_x
    else:
        ratio_y = random.uniform(1, max_ratio)

    oh, ow = int(h * ratio_y), int(w * ratio_x)
    off_y = random.randint(0, oh - h)
    off_x = random.randint(0, ow - w)

    dst = np.full(shape=(oh, ow, c), fill_value=fill, dtype=img.dtype)

    dst[off_y:off_y + h, off_x:off_x + w, :] = img

    # correct bbox
    bbox[:, :2] += (off_x, off_y)
    bbox[:, 2:4] += (off_x, off_y)

    return dst, bbox

def process_box(boxes, labels, img_size, class_num, anchors):
    anchors_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]

    # convert boxes form:
    # shape: [N, 2]
    # (x_center, y_center)
    box_centers = (boxes[:, 0:2] + boxes[:, 2:4]) / 2
    # (width, height)
    box_sizes = boxes[:, 2:4] - boxes[:, 0:2]

    # [13, 13, 3, 5+num_class+1] `5` means coords and labels. `1` means mix up weight. 
    y_true_13 = np.zeros((img_size[1] // 32, img_size[0] // 32, 3, 6 + class_num), np.float32)
    y_true_26 = np.zeros((img_size[1] // 16, img_size[0] // 16, 3, 6 + class_num), np.float32)
    y_true_52 = np.zeros((img_size[1] // 8, img_size[0] // 8, 3, 6 + class_num), np.float32)

    # mix up weight default to 1.
    y_true_13[..., -1] = 1.
    y_true_26[..., -1] = 1.
    y_true_52[..., -1] = 1.

    y_true = [y_true_13, y_true_26, y_true_52]

    # [N, 1, 2]
    box_sizes = np.expand_dims(box_sizes, 1)
    # broadcast tricks
    # [N, 1, 2] & [9, 2] ==> [N, 9, 2]
    mins = np.maximum(- box_sizes / 2, - anchors / 2)
    maxs = np.minimum(box_sizes / 2, anchors / 2)
    # [N, 9, 2]
    whs = maxs - mins

    # [N, 9]
    iou = (whs[:, :, 0] * whs[:, :, 1]) / (
                box_sizes[:, :, 0] * box_sizes[:, :, 1] + anchors[:, 0] * anchors[:, 1] - whs[:, :, 0] * whs[:, :,
                                                                                                         1] + 1e-10)
    # [N]
    best_match_idx = np.argmax(iou, axis=1)

    ratio_dict = {1.: 8., 2.: 16., 3.: 32.}
    for i, idx in enumerate(best_match_idx):
        # idx: 0,1,2 ==> 2; 3,4,5 ==> 1; 6,7,8 ==> 0
        feature_map_group = 2 - idx // 3
        # scale ratio: 0,1,2 ==> 8; 3,4,5 ==> 16; 6,7,8 ==> 32
        ratio = ratio_dict[np.ceil((idx + 1) / 3.)]
        x = int(np.floor(box_centers[i, 0] / ratio))
        y = int(np.floor(box_centers[i, 1] / ratio))
        k = anchors_mask[feature_map_group].index(idx)
        c = labels[i]
        # print(feature_map_group, '|', y,x,k,c)

        y_true[feature_map_group][y, x, k, :2] = box_centers[i]
        y_true[feature_map_group][y, x, k, 2:4] = box_sizes[i]
        y_true[feature_map_group][y, x, k, 4] = 1.
        y_true[feature_map_group][y, x, k, 5 + c] = 1.
        y_true[feature_map_group][y, x, k, -1] = boxes[i, -1]

    return y_true_13, y_true_26, y_true_52


def parse_data(data, class_num, img_size, anchors, is_training, letterbox_resize):
    
    img_idx, encoded_img, boxes, labels, _, _ = parse_tfrecord(data)
    img = cv2.imdecode(encoded_img, cv2.IMREAD_COLOR)
    boxes = np.concatenate((boxes, np.full(shape=(boxes.shape[0], 1), fill_value=1., dtype=np.float32)), axis=-1)

    ## I erased mix-up method here

    if is_training:
        # random color distortion
        img = random_color_distort(img)

        # random expansion with prob 0.5
        if np.random.uniform(0, 1) > 0.5:
            img, boxes = random_expand(img, boxes, 4)

        # random cropping
        h, w, _ = img.shape
        boxes, crop = random_crop_with_constraints(boxes, (w, h))
        x0, y0, w, h = crop
        img = img[y0: y0+h, x0: x0+w]

        # resize with random interpolation
        h, w, _ = img.shape
        interp = np.random.randint(0, 5)
        img, boxes = resize_with_bbox(img, boxes, img_size[0], img_size[1], interp=interp, letterbox=letterbox_resize)

        # random horizontal flip
        h, w, _ = img.shape
        img, boxes = random_flip(img, boxes, px=0.5)
    else:
        img, boxes = resize_with_bbox(img, boxes, img_size[0], img_size[1], interp=1, letterbox=letterbox_resize)

    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB).astype(np.float32)

    # the input of yolo_v3 should be in range 0~1
    img = img / 255.

    y_true_13, y_true_26, y_true_52 = process_box(boxes, labels, img_size, class_num, anchors)

    return img_idx, img, y_true_13, y_true_26, y_true_52


def get_batch_data(records, class_num, img_size, anchors, is_training, multi_scale=False, mix_up=False, letterbox_resize=True, interval=10):
    global iter_cnt

    # multi_scale training
    if multi_scale and is_training:
        random.seed(iter_cnt // interval)
        random_img_size = [[x * 32, x * 32] for x in range(10, 20)]
        img_size = random.sample(random_img_size, 1)[0]
    iter_cnt += 1

    img_idx_batch, img_batch, y_true_13_batch, y_true_26_batch, y_true_52_batch = [], [], [], [], []

    # deleted mix up strategy
    
    for data in records:
        img_idx, img, y_true_13, y_true_26, y_true_52 = parse_data(data, class_num, img_size, anchors, is_training, letterbox_resize)

        img_idx_batch.append(img_idx)
        img_batch.append(img)
        y_true_13_batch.append(y_true_13)
        y_true_26_batch.append(y_true_26)
        y_true_52_batch.append(y_true_52)

    img_idx_batch, img_batch, y_true_13_batch, y_true_26_batch, y_true_52_batch = np.asarray(img_idx_batch, np.int64), np.asarray(img_batch), np.asarray(y_true_13_batch), np.asarray(y_true_26_batch), np.asarray(y_true_52_batch)

    return img_idx_batch, img_batch, y_true_13_batch, y_true_26_batch, y_true_52_batch