Question about implementation of FB-OCC

MCG-NJU / SparseOcc

[ECCV 2024] Fully Sparse 3D Occupancy Prediction & RayIoU Evaluation Metric

https://arxiv.org/abs/2312.17118

Apache License 2.0

277 stars 22 forks source link

Question about implementation of FB-OCC #26

Closed LeeKeyu closed 5 months ago

LeeKeyu commented 6 months ago

Hi, thanks for your great and impressive work! I would like to ask about the implementation of FB-OCC in your paper. The FPS is about 10 Hz as reported in your paper, while I observed that the FPS of FB-OCC implemented using the same A100 GPU was about 2 Hz (for example in this paper OctreeOcc. Could you please elaborate more on this discrepancy? Your advice will be of great help to me and other potential readers. Thanks in advance!

afterthat97 commented 6 months ago

For FBOcc, we use the official config https://github.com/NVlabs/FB-BEV/blob/main/occupancy_configs/fb_occ/fbocc-r50-cbgs_depth_16f_16x4_20e.py and the provided weights to reproduce the results

Looks like OctreeOcc modified the setting? The input image size is set to 900×1600 and ResNet101-DCN is used as the image backbone, according to their paper.

LeeKeyu commented 6 months ago

May I ask how many GPUs did you use for evaluation? Thanks so much!

afterthat97 commented 6 months ago

For evaluation, single gpu is enough

LeeKeyu commented 5 months ago

I see. Thank you for your answer!