Per-scene mAP computation

Dear authors,

Thanks for this great work.

I would like to ask about more details of the per-scene mAP metrics.

In COCO-style mAP computation, all detections, i.e. fused 3DBBs, need to be sorted according to their confidences to draw the precision-recall curve.
In this method BoxerNet predicts a score for each monocular box, and then the score for the fused 3DBB is a weighted-average of all the scores in the cluster.

https://github.com/facebookresearch/boxer/blob/1f86542dc342a4b1d474c87c97c5d1d6566d9148/utils/fuse_3d_boxes.py#L655-L658

Is this score used to sort the detections?

Thanks a lot.

Read utils/fuse_3d_boxes.py around lines 655-658 and trace how the fused 3DBB score is used in per-scene mAP evaluation. Confirm the detection-ordering rule and record the explanation in the relevant project documentation or issue response; done means the COCO-style sorting behavior is unambiguous.

Written by the indexing model from the issue text.