ldkong1205 / LaserMix

[CVPR 2023 Highlight] LaserMix for Semi-Supervised LiDAR Semantic Segmentation
https://ldkong.com/LaserMix
Apache License 2.0
272 stars 17 forks source link

Training config for the nuScenes and ScribbleKITTI dataset. #13

Closed azhuantou closed 8 months ago

azhuantou commented 8 months ago

Hi, thanks for your nice work! Do you have any plan to release the config file of the lasermix for the nuScenes and ScribbleKITTI dataset?

ldkong1205 commented 8 months ago

Hi @azhuantou, thanks for reminding!

The config files are now available under the configs folder.

azhuantou commented 8 months ago

Thanks for your reply! @ldkong1205 But the corresponding files (e.g., semi_nuscenes_seg.py) are not updeted.
Could you please upload the missing files?

ldkong1205 commented 8 months ago

Hi @azhuantou,

Sorry for the confusion. The missing files are now uploaded to the codebase. Could you try and let me know if they work?

By the way, for the ScribbleKITTI dataset, since it shares the same data with SemanticKITTI (the only difference is the labels), you can just adopt the scripts from SemanticKITTI and modify the file paths.

azhuantou commented 8 months ago

Thanks for your timely reply! @ldkong1205 The corresponding pkl files (e.g., nuscenes_infos_train.20.pkl) for the nuscenes data are missing, and I can not find in https://drive.google.com/drive/folders/1PInw2Wvt-vgNzOxlSd2EiDANrTsWV7w1.

ldkong1205 commented 8 months ago

Hi @azhuantou, the nuscenes_infos_train.pkl has a relatively large size. We recommend generating it locally using generate_nus_pkl.py.

After this, you can use generate_nus_semi_pkl.py to generate the corresponding files for semi-supervised learning. These .pkl file will have a similar format as those used for SemanticKITTI/ScribbleKITTI.

azhuantou commented 8 months ago

Thank you for your patient responses. @ldkong1205 Following your guidance, I have successfully generated the pkl files (nuscenes_infos_train.pkl and nuscenes_infos_val.pkl with a size of 5521KB and 1181KB, respectively). Is that right? More important, frame indices for different ratios are not provided (e.g., nuscenes_1pct.txt), and this file might be necessary for reproducing your results. Looking forward to your reply!

ldkong1205 commented 8 months ago

Hi @azhuantou, the sizes of the original nuscenes_infos_train.pkl and nuscenes_infos_val.pkl files are listed as follows:

File Size
nuscenes_infos_train.pkl 469 MB
nuscenes_infos_val.pkl 98 MB

This is because the original nuScenes .pkl files contain information not only for lidarseg but also for 3D detection and multi-view images.

We have managed to generate .pkl files for our lidarseg task only. You can download them from this Google Drive. The dataset structure can be found at DATA_PREPARE.md.

The updated files' sizes are as follows:

File Size
nuscenes_infos_train.pkl 5.7 MB
nuscenes_infos_val.pkl 1.2 MB
nuscenes_infos_train.1.pkl 56 KB
nuscenes_infos_train.1-unlabeled.pkl 5.6 MB
nuscenes_infos_train.10.pkl 565 KB
nuscenes_infos_train.10-unlabeled.pkl 5.1 MB
nuscenes_infos_train.20.pkl 1.1 MB
nuscenes_infos_train.20-unlabeled.pkl 4.5 MB
nuscenes_infos_train.50.pkl 2.8 MB
nuscenes_infos_train.50-unlabeled.pkl 2.8 MB
azhuantou commented 8 months ago

Thank you so much! I will try the next step.

ldkong1205 commented 8 months ago

Thank you so much! I will try the next step.

Sure! Feel free to reach out again if you have other questions.

ldkong1205 commented 8 months ago

Hi @azhuantou, I attached one log file below for you to check. Without tuning the default parameters, you should at least reach the reported scores with <1% mIoU variance. Let us know if you find any problems with your reproduction. Thanks!

lasermix_cy3d_4xb4_semi_nuscenes_10.log