Official code for CFNet
Center Focusing Network for Real-Time LiDAR Panoptic Segmentation, Xiaoyan Li, Gang Zhang, Boyue Wang, Yongli Hu, Baocai Yin. (https://openaccess.thecvf.com/content/CVPR2023/papers/Li_Center_Focusing_Network_for_Real-Time_LiDAR_Panoptic_Segmentation_CVPR_2023_paper.pdf) Accepted by CVPR2023
- [2023-02-24] CFNet is accepted by CVPR 2023
- [2022-11-17] CFNet achieves the 63.4 PQ and 68.3 mIoU on the SemanticKITTI LiDAR Panoptic Segmentation Benchmark with the inference latency of 43.5 ms on a single NVIDIA RTX 3090 GPU.
CUDA>=11.1
Pytorch>=1.10.0
PyYAML>=6.0.0
scipy
pip3 install -r requirements.txtcd pytorch_lib
python setup.py installPlease download the SemanticKITTI dataset to the folder ./data and the structure of the folder should look like:
./data
├── SemanticKITTI
├── ...
└── dataset/
├──sequences
├── 00/
│ ├── velodyne/
| | ├── 000000.bin
| | ├── 000001.bin
| | └── ...
│ └── labels/
| ├── 000000.label
| ├── 000001.label
| └── ...
├── 08/ # for validation
├── 11/ # 11-21 for testing
└── 21/
└── ...
And download the object bank on the SemanticKITTI to the folder ./data and the structure of the folder should look like:
./data
├── object_bank_semkitti
├── bicycle
├── bicyclist
├── car
├── motorcycle
├── motorcyclist
├── other-vehicle
├── person
├── truck
torchrun --nproc_per_node=8 train.py --config config/semantickitti/config_mvfcev2ctx_adam_wce_lossv2_single.py --precision "fp32"torchrun --nproc_per_node=8 evaluate.py --config config/semantickitti/config_mvfcev2ctx_adam_wce_lossv2_single.py --precision "fp32" --resume_ckpt ${model_path}@inproceedings{licfnet2023,
author={Li, Xiaoyan and Zhang, Gang and Wang, Boyue and Hu, Yongli and Yin, Baocai},
booktitle={2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
title={Center Focusing Network for Real-Time LiDAR Panoptic Segmentation},
year={2023},
volume={},
number={},
pages={13425-13434},
doi={10.1109/CVPR52729.2023.01290}
}