`bbox` is set as `[0, 0, 50, 50]` by force in evaluate ？

[OuYaozhong]

Hi, OpenPCDet Team,

I notices that, the function transform_annotations_to_kitti_format inside the official code pcdet/datasets/kitti/kitti_utils.py,

it sets the bbox attribute of both gt and dt as [0, 0, 50, 50] by force
ref

anno['bbox'] = np.zeros((len(anno['name']), 4))
anno['bbox'][:, 2:4] = 50  # [0, 0, 50, 50]

and

ref

kitti_utils.transform_annotations_to_kitti_format(eval_det_annos, map_name_to_kitti=map_name_to_kitti)
kitti_utils.transform_annotations_to_kitti_format(
    eval_gt_annos, map_name_to_kitti=map_name_to_kitti,
    info_with_fakelidar=self.dataset_cfg.get('INFO_WITH_FAKELIDAR', False)
)

I can not find any issue reported this problem.

And when I debug at the position_1, the position_2 , and position_3, where calculates the difficulty and tp, fp, fn, similarity, thresholds, I found that all the height which calculated from bbox attribute are all 50s, or bbox related variables are all [0, 0, 50, 50].

Can the team member or some one familiar with these can give some explanation ?

[OuYaozhong]

I found some relative contents (#253 (comment)), but seems no one answer the question.

For my temporary understanding, It seems that the bbox will not affect the bev and 3d result, since these two is related to the xxx_lidar attributes, and the fake bbox only affect the bbox results.

And the reason why set bbox as [0, 0, 50, 50] is to make it will not be filtered by the difficulty, since KITTI will filter the height smaller than 25.

And the difficulty definition for custom dataset need to be defined by self, so that the occluded and truncated are both set as 0 as a template and need to be adjust by self, which can avoid being filtered by difficulty.

Is this right ? It is a best understanding for now from me.