search

facebookresearch / ContrastiveSceneContexts

Code for CVPR 2021 oral paper "Exploring Data-Efficient 3D Scene Understanding with Contrastive Scene Contexts"
MIT License
224 stars 28 forks source link

question about NMS in instance segmentation #39

Closed Dingry closed 2 years ago

Dingry commented 2 years ago

Hi, thanks for your code release. I have a question about the nms code here. It seems that it only removes some proposals predicted in [10, 12, 16] classes, not like a regular NMS where the proposal score is utilized for proposal ranking and the iou overlap between proposals are computed to remove redundant proposals?

Sekunde commented 2 years ago

Thanks for raising it up. That is due to debugging purpose and you can just comment it out.

Dingry commented 2 years ago

Thanks for you reply. Do you mean that the NMS is not needed in final inference?

Sekunde commented 2 years ago

Hi, NMS (this implementation is just a selection between the two sets of clustering results, not really NMS)is only used when enabling dual set clustering, and we do not use the dual set clustering for the paper setup; it is only used when submitting to the ScanNet benchmark.

Dingry commented 2 years ago

Thanks for your reply. My questions are addressed.