CheXpert bbox annotations scaling

[uizdd]

I have trained the model with the CheXpert dataset (440Gb version). Unfortunately, the generated masks are all scaled incorrectly.
https://we.tl/t-BejaX6Plgg (The generated masks seem to be too large by a factor of about 1.3.)

For the NIH Chest X-ray14 dataset, everything works fine.

My understanding is that the images and bounding boxes are always transformed together. So it is not clear to me how such an error can occur.

[uizdd]

Due to the annotation format differing from the NIH Chest X-ray14 dataset, I suggest the following changes if you are using the CheXpert dataset:

net_utils.py

    #factor = float(config['img_size']) / 1024.           
    factor_h = float(config['img_size']) / 2320.
    factor_w = float(config['img_size']) / 2828.
    for b_i in range(len(config['Data_CLASSES'])):
        if config['Data_CLASSES'][b_i] in bbox_img_gt:
            for loc in bbox_img_gt[config['Data_CLASSES'][b_i]]:
                #bbox_img = cv2.rectangle(bbox_img, (int(loc[0]*factor), int(loc[1]*factor)), (int((loc[0]+loc[2])*factor), int((loc[1]+loc[3])*factor)), (palette[b_i+1][2],palette[b_i+1][1],palette[b_i+1][0]), 4)
                bbox_img = cv2.rectangle(bbox_img, (int(loc[0]*factor_w), int(loc[1]*factor_h)), (int(loc[2]*factor_w), int(loc[3]*factor_h)), (palette[b_i+1][2],palette[b_i+1][1],palette[b_i+1][0]), 4)
    cv2.imwrite(bbox_path, bbox_img)

custom_dataset.py

    class BBoxToMask(object):
        def __call__(self, bboxs, classes, img, im_w, im_h, name, using_crf):
            img = np.array(img)
            img = np.ascontiguousarray(img)
            mask = []
            for idx in range(len(classes)):
                class_gts = np.zeros((im_h, im_w))
                if classes[idx] in bboxs:
                    mask_h = 0
                    mask_w = 0
                    j = 0
                    for loc in bboxs[classes[idx]]:
                        #mask_h += loc[2]
                        mask_h += (loc[2] - loc[2])
                        #mask_w += loc[3]
                        mask_w += (loc[3] - loc[1])
                        #class_gts[int(loc[1]):int(loc[1]+loc[3]), int(loc[0]):int(loc[0]+loc[2])] = 1
                        class_gts[int(loc[1]):int(loc[3]), int(loc[0]):int(loc[2])] = 1
                        j += 1