search

tusen-ai / SST

Code for a series of work in LiDAR perception, including SST (CVPR 22), FSD (NeurIPS 22), FSD++ (TPAMI 23), FSDv2, and CTRL (ICCV 23, oral).
Apache License 2.0
801 stars 102 forks source link

The results of FSD on nuscenes or kitti #89

Closed Pedrojj97 closed 1 year ago

Pedrojj97 commented 1 year ago

Hi, have you ever tried the fsd model on nuscenes or kitti.

Abyssaledge commented 1 year ago

Someone has tried FSD on KITTI: https://github.com/tusen-ai/SST/issues/77 The design of the nuScenes version is slightly different due to its evaluation metric being quite different from the WOD. We will release it in a couple of months.

Pedrojj97 commented 1 year ago

Thanks for your quickly replay, Would you mind telling me what the slightly difference is? In my opinion, fsd depends on the result of segmentation. Therefore, whether it is not very friendly to the datasets consist of many categories like nuscenes.

Abyssaledge commented 1 year ago

It‘s a misunderstanding. FSD can well handle the Argoverse 2 dataset in our paper, which contains up to 30 classes. Differences are:

  1. nuScenes does not use IoU as a metric, so the label assignment in the GroupCorrection head should not use IoU.
  2. In WOD, a point cloud could belong to more than 1 category (Binary sigmoid CE loss as segmentation loss). In nus or Argo, we restrict a point only belonging to a single class, so we use softmax CE loss as segmentation loss and the group_sample function to sample fg.
Pedrojj97 commented 1 year ago

I got it, thanks again~