Skip to content

Latest commit

 

History

History
125 lines (80 loc) · 5.54 KB

README.md

File metadata and controls

125 lines (80 loc) · 5.54 KB

Revisiting Retentive Networks for Fast Range-View 3D LiDAR Semantic Segmentation

Simone Mosco, Daniel Fusaro, Wanmeng Li, Alberto Pretto

University of Padova

WACV 2026

Paper PDF HF Model YouTube Video

teaser

RangeRet is a lightweight LiDAR semantic segmentation approach that adapts Retentive Networks to range images using a Circular Retention (CiR) mechanism to efficiently model spatial continuty properties of range-view representation. RangeRet obtains SOTA results among range-view approaches on PandaSet and SemanticPOSS, and competitive results on SemanticKITTI, while achieving real-time performance.

News

  • 2026-02-25: Code and models are released.
  • 2025-09-05: RangeRet is accepted to WACV 2026.

Installation

Create an environment and install the required packages with pip install -r requirements.txt.

In practice, you just need common packages as torch, numpy, matplotlib, pyyaml, timm, tqdm.

Code has been tested with the following versions of torch:

  • torch 2.2.2 and CUDA 12.1
  • torch 2.4.0 and CUDA 12.4

Data Preparation

SemanticKITTI

Download SemanticKITTI dataset from the official website.

PandaSet

Download PandaSet dataset from the official website and raw Pandar64 LiDAR data from the link provided in this Github Issue.

Follow the instructions on this repo to convert the PandaSet dataset to the SemanticKITTI format.

SemanticPOSS

Download SemanticPOSS dataset from the official website.

Training

Run the following commands to train, specifying the dataset and log paths:

### SemanticKITTI
python train.py --dataset /path/to/semantickitti/ --data ./config/labels/semantic-kitti.yaml --config ./config/RangeRet-semantickitti.yaml --log ./log/kitti [--fp16]

### PandaSet
python train.py --dataset /path/to/pandaset/ --data ./config/labels/pandaset.yaml --config ./config/RangeRet-pandaset.yaml --log ./log/panda [--fp16]

### SemanticPOSS
python train.py --dataset /path/to/semanticposs/ --data ./config/labels/semantic-poss.yaml --config ./config/RangeRet-poss.yaml --log ./log/poss [--fp16]

Testing

Run the following scripts to infer on a specific dataset:

### SemanticKITTI
python infer.py --dataset /path/to/semantickitti/ --data ./config/labels/semantic-kitti.yaml --config ./config/RangeRet-semantickitti.yaml --model /path/to/model.pt --split valid/test --log /path/to/predictions --fp16 [--save]

### PandaSet
python infer.py --dataset /path/to/pandaset/ --data ./config/labels/pandaset.yaml --config ./config/RangeRet-pandaset.yaml --model /path/to/model.pt --split valid/test --log /path/to/predictions --fp16 [--save]

### SemanticPOSS
python infer.py --dataset /path/to/semanticposs/ --data ./config/labels/semantic-poss.yaml --config ./config/RangeRet-poss.yaml --model /path/to/model.pt --split valid/test --log /path/to/predictions --fp16 [--save]

### E.g. for SemanticKITTI and pretrained model rangeret-kitti-657
# python3 infer.py --dataset /semanticKITTI/ --data ./config/labels/semantic-kitti.yaml --config config/RangeRet-semantickitti.yaml --model ./rangeret-kitti-657.pt --split valid --fp16 [--save] --log ./out/kitti_results

Model Zoo

Pretrained models are available on Hugging Face.

Name Dataset Val mIoU (%) Test mIoU (%) Checkpoints
rangeret-kitti-657 SemanticKITTI 65.7 64.5 Download
rangeret-panda-600 PandaSet 66.7 60.0 Download
rangeret-poss-528 SemanticPOSS - 52.8 Download

Checklist

  • Update README
  • Release code
  • Upload models
  • Custom data guidelines

Citation

If you find our work useful please cite our paper:

@inproceedings{mosco2026revisiting,
  title={Revisiting Retentive Networks for Fast Range-View 3D LiDAR Semantic Segmentation},
  author={Mosco, Simone and Fusaro, Daniel and Li, Wanmeng and Pretto, Alberto},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={2499--2509},
  year={2026}
}

Acknowledgment

Thanks to these great repositories: RangeNet++, Torchscale, RMT, RangeViT, Rangeview-rgb-lidar-fusion