[拆轮子] PaddleDetection 中的预处理 NormalizeBox、 BboxXYXY2XYWH 和 Permute

NormalizeBox 和 NormalizeImage 很像，都是缩放到某个区间

NormalizeBox 就是把gt_bbox的位置缩放到 [0,1]

class NormalizeBox(BaseOperator):
    """Transform the bounding box's coornidates to [0,1]."""

    def __init__(self):
        super(NormalizeBox, self).__init__()

    def apply(self, sample, context):
        im = sample['image']
        gt_bbox = sample['gt_bbox']
        height, width, _ = im.shape
		
		# ------------ 缩放边界框 gt_bbox ------------
        for i in range(gt_bbox.shape[0]):
            gt_bbox[i][0] = gt_bbox[i][0] / width
            gt_bbox[i][1] = gt_bbox[i][1] / height
            gt_bbox[i][2] = gt_bbox[i][2] / width
            gt_bbox[i][3] = gt_bbox[i][3] / height
        sample['gt_bbox'] = gt_bbox
		
		# ------------ 缩放关键点部分 ------------
        if 'gt_keypoint' in sample.keys():
            gt_keypoint = sample['gt_keypoint']
			
			# 关键点的标注顺序 x1y1x2y2x3y3....
            for i in range(gt_keypoint.shape[1]):
                if i % 2:
                    gt_keypoint[:, i] = gt_keypoint[:, i] / height
                else:
                    gt_keypoint[:, i] = gt_keypoint[:, i] / width
            sample['gt_keypoint'] = gt_keypoint

        return sample

class BboxXYXY2XYWH(BaseOperator):
    """
    Convert bbox XYXY format to XYWH format.
    """

    def __init__(self):
        super(BboxXYXY2XYWH, self).__init__()

    def apply(self, sample, context=None):
        assert 'gt_bbox' in sample
        bbox = sample['gt_bbox']
        bbox[:, 2:4] = bbox[:, 2:4] - bbox[:, :2]       # 拿到宽高
        bbox[:, :2] = bbox[:, :2] + bbox[:, 2:4] / 2.   # 将左上角移动至中心
        sample['gt_bbox'] = bbox
        return sample

没传入该 apply 可视化：

im = sample['image']
img = (sample['image'] * [0.229, 0.224,0.225] + [0.485, 0.456, 0.406]) * 255
img = img.astype(int).astype("uint8")
h, w, _ = img.shape
gt_bboxes = sample['gt_bbox'] * [w, h, w, h]

gt_bboxes = gt_bboxes.astype(int)
x1, y1, x2, y2 = gt_bboxes[1]

xx = cv2.rectangle(img, (x1, y1), (x2, y2), 255, thickness=2, lineType=8)
cv2.imwrite("xxx.png", xx)

传出该 apply 可视化：

im = sample['image']
img = (sample['image'] * [0.229, 0.224,0.225] + [0.485, 0.456, 0.406]) * 255
img = img.astype(int).astype("uint8")
h, w, _ = img.shape
gt_bboxes = sample['gt_bbox'] * [w, h, w, h]

gt_bboxes = gt_bboxes.astype(int)
x_c, y_c, w, h = gt_bboxes[3]
x1, y1, x2, y2 = x_c-w//2, y_c-h//2, x_c+w//2, y_c+h//2

xx = cv2.rectangle(img, (x1, y1), (x2, y2), 255, thickness=2, lineType=8)
cv2.imwrite("xxx.png", xx)

注意此处的 sample['h'] sample['w'] 都是原图的尺寸

Permute 的话就很简单了，就是 HWC 转化为 CHW

class Permute(BaseOperator):
    def __init__(self):
        """
        Change the channel to be (C, H, W)
        """
        super(Permute, self).__init__()

    def apply(self, sample, context=None):
        im = sample['image']
        im = im.transpose((2, 0, 1))
        sample['image'] = im
        return sample