看代码的时候遇到的问题

[linhaibo123]

作者这个修改的代码很有用，帮了我大忙，但是还是有个地方没有看明白 在data_util.py 中，有个参数n,m。n设置为6，m为0.5。看代码在

def generate_rbox(im_size, text_polys, text_tags, training_mask, i, n, m):
    """
    生成mask图，白色部分是文本，黑色是北京
    :param im_size: 图像的h,w
    :param text_polys: 框的坐标
    :param text_tags: 标注文本框是否参与训练
    :return: 生成的mask图
    """
    h, w = im_size
    score_map = np.zeros((h, w), dtype=np.uint8)
    for poly, tag in zip(text_polys, text_tags):
        poly = poly.astype(np.int)
        r_i = 1 - (1 - m) * (n - i) / (n - 1)
        d_i = cv2.contourArea(poly) * (1 - r_i * r_i) / cv2.arcLength(poly, True)
        pco = pyclipper.PyclipperOffset()
        # pco.AddPath(pyclipper.scale_to_clipper(poly), pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
        # shrinked_poly = np.floor(np.array(pyclipper.scale_from_clipper(pco.Execute(-d_i)))).astype(np.int)
        pco.AddPath(poly, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
        shrinked_poly = np.array(pco.Execute(-d_i))
        cv2.fillPoly(score_map, shrinked_poly, 1)
        # 制作mask
        # rect = cv2.minAreaRect(shrinked_poly)
        # poly_h, poly_w = rect[1]
        # if min(poly_h, poly_w) < 10:
        #     cv2.fillPoly(training_mask, shrinked_poly, 0)
        if tag:
            cv2.fillPoly(training_mask, shrinked_poly, 0)
        # 闭运算填充内部小框
        # kernel = np.ones((3, 3), np.uint8)
        # score_map = cv2.morphologyEx(score_map, cv2.MORPH_CLOSE, kernel)
    return score_map, training_mask
def augmentation(im: np.ndarray, text_polys: np.ndarray, scales: np.ndarray, degrees: int, input_size: int) -> tuple:
    # the images are rescaled with ratio {0.5, 1.0, 2.0, 3.0} randomly
    im, text_polys = data_aug.random_scale(im, text_polys, scales)
    # the images are horizontally fliped and rotated in range [−10◦, 10◦] randomly
    if random.random() < 0.5:
        im, text_polys = data_aug.horizontal_flip(im, text_polys)
    if random.random() < 0.5:
        im, text_polys = data_aug.random_rotate_img_bbox(im, text_polys, degrees)
    # 640 × 640 random samples are cropped from the transformed images
    # im, text_polys = data_aug.random_crop_img_bboxes(im, text_polys)

    # im, text_polys = data_aug.resize(im, text_polys, input_size, keep_ratio=False)
    # im, text_polys = data_aug.random_crop_image_pse(im, text_polys, input_size)
    return im, text_polys
def image_label(im_fn: str, text_polys: np.ndarray, text_tags: list, n: int, m: float, input_size: int,
                defrees: int = 10,
                scales: np.ndarray = np.array([0.5, 1, 2.0, 3.0])) -> tuple:
    '''
    get image's corresponding matrix and ground truth
    return
    images [512, 512, 3]
    score  [128, 128, 1]
    geo    [128, 128, 5]
    mask   [128, 128, 1]
    '''
    im = cv2.imread(im_fn)
    im = cv2.cvtColor(im,cv2.COLOR_BGR2RGB)
    h, w, _ = im.shape
    # 检查越界
    text_polys = check_and_validate_polys(text_polys, (h, w))
    im, text_polys, = augmentation(im, text_polys, scales, defrees, input_size)
    h, w, _ = im.shape
    short_edge = min(h, w)
    if short_edge < input_size:
        # 保证短边 >= inputsize
        scale = input_size / short_edge
        im = cv2.resize(im, dsize=None, fx=scale, fy=scale)
        text_polys *= scale
    # # normal images
    # im = im.astype(np.float32)
    # im /= 255.0
    # im -= np.array((0.485, 0.456, 0.406))
    # im /= np.array((0.229, 0.224, 0.225))
    h, w, _ = im.shape
    training_mask = np.ones((h, w), dtype=np.uint8)
    score_maps = []
    for i in range(1, n + 1):
        # s1->sn,由小到大
        score_map, training_mask = generate_rbox((h, w), text_polys, text_tags, training_mask, i, n, m)
        score_maps.append(score_map)
    score_maps = np.array(score_maps, dtype=np.float32)
    imgs = data_aug.random_crop_author([im, score_maps.transpose((1, 2, 0)),training_mask], (input_size, input_size))
    return imgs[0], imgs[1].transpose((2, 0, 1)), imgs[2]#im,score_maps,training_mask#

中有这么一段，

 poly = poly.astype(np.int)
        r_i = 1 - (1 - m) * (n - i) / (n - 1)
        d_i = cv2.contourArea(poly) * (1 - r_i * r_i) / cv2.arcLength(poly, True)
        pco = pyclipper.PyclipperOffset()
        # pco.AddPath(pyclipper.scale_to_clipper(poly), pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
        # shrinked_poly = np.floor(np.array(pyclipper.scale_from_clipper(pco.Execute(-d_i)))).astype(np.int)
        pco.AddPath(poly, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
        shrinked_poly = np.array(pco.Execute(-d_i))
        cv2.fillPoly(score_map, shrinked_poly, 1)

这段应该是对poly进行了缩放，但是为什么要这么计算和设置呢？目的是什么呢？能帮忙解答下吗？谢谢。

[WenmuZhou]

参考论文