Meta-RangeSeg

Song Wang, Jianke Zhu*, Ruixiang Zhang

This is the official implementation of Meta-RangeSeg: LiDAR Sequence Semantic Segmentation Using Multiple Feature Aggregation [Paper] [Video].

	Prediction	Groud Truth
Perspective View
Bird's-Eye View

Demo

Model Zoo

Model	Task	mIoU(paper) [on test set]	mIoU(reprod.) [on test set]	Results
Meta-RangeSeg	multiple scans semantic segmentation	49.5	49.7	valid_pred test_pred
Meta-RangeSeg	single scan semantic segmentation	61.0	60.3	valid_pred test_pred

Data Preparation

SemanticKITTI download

Please download the original SemanticKITTI dataset from the official website.

Residual image generation

For residual image generation, we provide an online version but adopt the offline one in the actual training. Please refer to LiDAR-MOS for more details. Thanks for their great work!

Testing Pretrained Models

You can run the following command to test the performance of Meta-RangeSeg:

cd ./train/tasks/semantic
python infer.py -d ./data/semantic_kitti/dataset -m ../../../logs

Training

To train the model from scratch, you can run:

CUDA_VISIBLE_DEVICES=0,1 python train.py -d ./data/semantic_kitti/dataset -ac ../../../meta_rangeseg.yml

Acknowledgment

This project is heavily based on SalsaNext and LiDAR-MOS. RangeDet and FIDNet are also excellent range-based models, which help us a lot.

Citations

@article{wang2022meta,
  title={Meta-RangeSeg: LiDAR Sequence Semantic Segmentation Using Multiple Feature Aggregation},
  author={Wang, Song and Zhu, Jianke and Zhang, Ruixiang},
  journal={IEEE Robotics and Automation Letters},
  volume={7},
  number={4},
  pages={9739--9746},
  year={2022},
  publisher={IEEE}
}

songw-zju / Meta-RangeSeg

readme