The results of using seed label visualization when reproducing the paper are not quite accurate

YurongYou / MODEST

Code release for "Learning to Detect Mobile Objects from LiDAR Scans Without Labels" [CVPR 2022]

MIT License

86 stars 11 forks source link

The results of using seed label visualization when reproducing the paper are not quite accurate #9

Open Truring opened 7 months ago

Truring commented 7 months ago

I used OpenPCDet's visualization method to visualize seed labels, but it didn't look quite right. The 3D boxes are floating on the ground and cannot properly wrap around obstacles. The point cloud I used is "*. bin" files in data_preprocessing/nuscenes/NUSCENESS_KITTI-FORMAT/training/velodyne, and the seed label is generated according to readme into generatecluster_mask/intermediate_results/nusc_bbox_pp_score_fw_20_r0.3.

Truring commented 6 months ago

Sorry for bothering you. Can you help me with this issue? @YurongYou @katieluo88

Harry710887048 commented 5 months ago

I'm encountering the same issue. Could it be a problem during the file generation, or is there an error in the coordinate transformation?

YurongYou commented 5 months ago

Hi Truring and Harry710887048, thanks for your interest! Looks like it is a file miss match issue, it could be something wrong when setting the file paths. Could you check the following?

Could you check the number of .txt files generated in the intermediate_results/nusc_bbox_pp_score_fw_20_r0.3 is equal to 3985 (training set size)?
Could you paint the point cloud with corresponding pp score? (intermediate_results/nusc_pp_score_fw_30_r0.3/)