def images_and_boxes_preprocessing()

in lib/datasets/data_input_helper.py [0:0]

• 44 lines of code
• 18 McCabe index (conditional complexity)

def images_and_boxes_preprocessing(
        imgs, split, crop_size, spatial_shift_pos, boxes=None):

    height, width, _ = imgs[0].shape

    if boxes is not None:
        boxes[:, [0, 2]] *= width
        boxes[:, [1, 3]] *= height

        boxes = imgproc.clip_boxes_to_image(boxes, height, width)

    # Now the image is in HWC, BGR format
    if split == 1:  # "train"
        imgs, boxes = imgproc.random_short_side_scale_jitter_list(
            imgs,
            min_size=cfg.TRAIN.JITTER_SCALES[0],
            max_size=cfg.TRAIN.JITTER_SCALES[1],
            boxes=boxes,
        )
        imgs, boxes = imgproc.random_crop_list(
            imgs, crop_size, order='HWC', boxes=boxes)

        # random flip
        imgs, boxes = imgproc.horizontal_flip_list(
            0.5, imgs, order='HWC', boxes=boxes)
    else:
        # Short side to cfg.TEST_SCALE. Non-local and STRG uses 256.
        imgs = [imgproc.scale(cfg.TEST.SCALE, img) for img in imgs]
        if boxes is not None:
            boxes = imgproc.scale_boxes(cfg.TEST.SCALE, boxes, height, width)

        if cfg.AVA.FORCE_TEST_FLIP and cfg.DATASET == 'ava':
            imgs, boxes = imgproc.horizontal_flip_list(
                0.5, imgs, order='HWC', boxes=boxes,
                force_flip=True)

        # For the short side we do center crop.
        imgs, boxes = imgproc.spatial_shift_crop_list(
            crop_size, imgs, spatial_shift_pos, boxes=boxes)

    # Convert image to CHW keeping BGR order
    imgs = [imgproc.HWC2CHW(img) for img in imgs]

    # image [0, 255] -> [0, 1]
    imgs = [img / 255.0 for img in imgs]

    imgs = [np.ascontiguousarray(
        img.reshape((3, crop_size, crop_size))).astype(np.float32)
        for img in imgs]

    # do color augmentation (after divided by 255.0)
    if cfg.TRAIN.USE_COLOR_AUGMENTATION and split == 1:
        imgs = color_augmentation_list(imgs)

    # now, normalize by mean and std
    imgs = [imgproc.color_normalization(img, DATA_MEAN, DATA_STD)
            for img in imgs]

    # 3, 224, 224 -> 3, 32, 224, 224
    imgs = np.concatenate(
        [np.expand_dims(img, axis=1) for img in imgs], axis=1)

    # Kinetics pre-training uses RGB!!
    if not cfg.MODEL.USE_BGR:
        # BGR to RGB.
        imgs = imgs[::-1, ...]

    if boxes is not None:
        boxes = imgproc.clip_boxes_to_image(boxes, crop_size, crop_size)
    return imgs, boxes