search

ch-sa / labelCloud

A lightweight tool for labeling 3D bounding boxes in point clouds.
https://ch-sa.github.io/labelCloud/
GNU General Public License v3.0
601 stars 103 forks source link

Is there any evaluation metrics for kitti untransformed kind dataset on 3D object detection? #156

Closed NyquistBodeTu closed 7 months ago

NyquistBodeTu commented 9 months ago

Hi, it's my great pleasure that the labelcloud is the wonderful tool to make KITTI untransformed kind dataset. By modifying some codes, it is easy to train 3D detection neural network. However, the evaluation metrics of KITTI dataset is different from ours. So I need to find a new evaluation metric for our dataset, just LiDAR data, without Images. Do you have any suggestions? Looking forward to your reply.

ch-sa commented 7 months ago

Usually the Intersection over Union (IoU) metric is used for evaluation.

There is another library bbox here on github that also has some support for evaluating 3D bounding boxes: https://github.com/varunagrawal/bbox/blob/master/bbox%2Fmetrics.py#L139

However, it does only consider the yaw orientation (top down).

Comparing fully rotated bounding boxes is not so trivial.