rot3d_range Causes Incorrect 3D Supervision？

[xlb8]

In the code, rot3d_range is used in the config to control the transformation of the 3D box, and its corresponding rotate_3d is added to lidar2img. However, your image does not have a corresponding rotate_3d transformation, which is a mismatch.
In other words, the depth supervision in the code is completely wrong.

    def __call__(self, results):
        angle = results["aug_config"]["rotate_3d"]
        rot_cos = np.cos(angle)
        rot_sin = np.sin(angle)

        rot_mat = np.array(
            [
                [rot_cos, -rot_sin, 0, 0],
                [rot_sin, rot_cos, 0, 0],
                [0, 0, 1, 0],
                [0, 0, 0, 1],
            ]
        )
        rot_mat_inv = np.linalg.inv(rot_mat)

        num_view = len(results["lidar2img"])
        for view in range(num_view):
            results["lidar2img"][view] = (
                results["lidar2img"][view] @ rot_mat_inv 
            )   # do the 3d transformation here
        if "lidar2global" in results:
            results["lidar2global"] = results["lidar2global"] @ rot_mat_inv
        if "gt_bboxes_3d" in results:
            results["gt_bboxes_3d"] = self.box_rotate(
                results["gt_bboxes_3d"], angle
            )
        return results

    def _img_transform(self, img, aug_configs):
        H, W = img.shape[:2]
        resize = aug_configs.get("resize", 1)
        resize_dims = (int(W * resize), int(H * resize))
        crop = aug_configs.get("crop", [0, 0, *resize_dims])
        flip = aug_configs.get("flip", False)
        rotate = aug_configs.get("rotate", 0)
        # there is no rorate_3d op for img aug