Why do you filter ground truth data on the condition h > 0 and w > 0 in the JointDataset.

[janthmueller]

Hi, you have within your implementation the condition h > 0 and w > 0 to filter every object out which does not satisfy it. But it seems only possible to hit a 0 on w or h if the down_ratio is larger than the image shape or if the labels from the fetched dataset have <= 0 width or height.

Have you observed such values in one of your used datasets or do you use preprocessing which can cause bbox w and h values to be <= 0?

Plus you seem to use reg_mask to "register" valid objects within the dataset. Is there a reason you called it reg_mask? Do you not use it for loss calculations on wh and re-id too?

        output_h = imgs.shape[1] // self.opt.down_ratio
        output_w = imgs.shape[2] // self.opt.down_ratio
        num_classes = self.num_classes
        num_objs = labels.shape[0]
        hm = np.zeros((num_classes, output_h, output_w), dtype=np.float32)
        if self.opt.ltrb:
            wh = np.zeros((self.max_objs, 4), dtype=np.float32)
        else:
            wh = np.zeros((self.max_objs, 2), dtype=np.float32)
        reg = np.zeros((self.max_objs, 2), dtype=np.float32)
        ind = np.zeros((self.max_objs, ), dtype=np.int64)
        reg_mask = np.zeros((self.max_objs, ), dtype=np.uint8)
        ids = np.zeros((self.max_objs, ), dtype=np.int64)
        bbox_xys = np.zeros((self.max_objs, 4), dtype=np.float32)

        draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else draw_umich_gaussian
        for k in range(min(num_objs, self.max_objs)):
            label = labels[k]
            bbox = label[2:]
            cls_id = int(label[0])
            bbox[[0, 2]] = bbox[[0, 2]] * output_w
            bbox[[1, 3]] = bbox[[1, 3]] * output_h
            bbox_amodal = copy.deepcopy(bbox)
            bbox_amodal[0] = bbox_amodal[0] - bbox_amodal[2] / 2.
            bbox_amodal[1] = bbox_amodal[1] - bbox_amodal[3] / 2.
            bbox_amodal[2] = bbox_amodal[0] + bbox_amodal[2]
            bbox_amodal[3] = bbox_amodal[1] + bbox_amodal[3]
            bbox[0] = np.clip(bbox[0], 0, output_w - 1)
            bbox[1] = np.clip(bbox[1], 0, output_h - 1)
            h = bbox[3]
            w = bbox[2]

            bbox_xy = copy.deepcopy(bbox)
            bbox_xy[0] = bbox_xy[0] - bbox_xy[2] / 2
            bbox_xy[1] = bbox_xy[1] - bbox_xy[3] / 2
            bbox_xy[2] = bbox_xy[0] + bbox_xy[2]
            bbox_xy[3] = bbox_xy[1] + bbox_xy[3]

            if h > 0 and w > 0:
                radius = gaussian_radius((math.ceil(h), math.ceil(w)))
                radius = max(0, int(radius))
                radius = 6 if self.opt.mse_loss else radius
                #radius = max(1, int(radius)) if self.opt.mse_loss else radius

[janthmueller]

I observed values of 0 in width/height in coco ds