CPGNet: Cascade Point-Grid Fusion Network for Real-Time LiDAR Semantic Segmentation

• This is Non-Official implementation of CPGNet. Just a simple try and just simple reproduction based on SMVF. :joy::joy::joy:
• Here is the official Repo.

New Fix:

• Remove transformation consistency loss due to its large training burden. (~2× GPU consumption & ~2× training time boost)
• CutMiX data augmentation now available.
• CosineAnnealingWarmUpRestarts look better than StepLR.
• Now, there is not much performance gap between wce and ohem.

Environment Setup

Please refer to SMVF repo. Note: Make sure deep_point is installed.

Prepare Data

Download SemanticKITTI from official web. Download Object_Bank from SMVF for CutMix.

Training

### Multi-gpus ###
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 train.py --config config/wce.py

### Single-gpu ###
CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 train.py --config config/wce.py

Evaluation

### Multi-gpus ###
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 evaluate.py --config config/wce.py --start_epoch 0 --end_epoch 49

### Single-gpu ###
CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 evaluate.py --config config/wce.py --start_epoch 0 --end_epoch 49

Find best epoch

python find_best_metric.py --name wce

Pretrained Models and Logs

Models have been uploaded to this Google Drive folder.

Note:

• Our model is trained using 1/4 of the data. See here of code.
• We did not use TTA, which actually slightly improved performance by about 1.0%. If want to enable it, comment here and here.

Below are known issues listed:

• Over-fitting on the validation set, the model with 64.9mIoU on the validation set achieved only 61.8mIoU on the CodaLab online test set. (Probably because we only use 1/4 of the data for training)

Citation

It should be considered to cite:

@inproceedings{li2022cpgnet,
  title={CPGNet: Cascade Point-Grid Fusion Network for Real-Time LiDAR Semantic Segmentation},
  author={Li, Xiaoyan and Zhang, Gang and Pan, Hongyu and Wang, Zhenhua},
  booktitle={2022 IEEE International Conference on Robotics and Automation (ICRA)},
  year={2022},
  organization={IEEE}
}