<p align="center"> <img src="assets/IAL-Logo.png" width="200"/> <p>

🔥How Do Images Align and Complement LiDAR? Towards a Harmonized Multi-modal 3D Panoptic Segmentation

🥳Accepted at ICML 2025🥳

IAL is a novel multi-modal framework for LiDAR-camera 3D panoptic segmentation. To address misalignment and weak fusion, IAL designs a synchronized augmentation (PieAug) and a geometric-guided fusion module (GTF) to align LiDAR and image features. Moreover, it leverages complementary priors from both modalities as queries through PQG for stronger instance prediction. On nuScenes, IAL achieves 82.3% PQ, and on SemanticKITTI, it obtains 63.1% PQ, significantly surpassing previous methods.

Demo

Framework Diagram

Preparation

Environments

Python == 3.8
CUDA == 11.1
pytorch == 1.10.1
mmcv == 2.0.0rc4
mmdet == 3.0.0
mmdet3d == 1.1.0
torch-scatter == 2.0.9
nms_lib

*Please see install.sh for more details.

Data Structure

Follow the mmdet3d to process the nuScenes dataset.

data/nuscenes_full/
├── gsam # save 2d proposals
├── lidarseg
├── maps
├── nuscenes_infos_test.pkl
├── nuscenes_infos_train.pkl
├── nuscenes_infos_val.pkl
├── panoptic
├── samples
├── sweeps
├── v1.0-trainval

You can generate *.pkl by excuting

python tools/create_data.py nuscenes --root-path data/nuscenes_full --out-dir data/nuscenes_full --extra-tag nuscenes

2D proposal Generation

# setup for Grounding-DINO
cd tools
git clone https://github.com/IDEA-Research/GroundingDINO.git
cd GroundingDINO
pip install -e .
# infer & save results
3-infer-gsam.sh

*Checkpoints and 2D preprocessing data are available huggingface and OneDrive.

Train & Inference

# train
sh 1-train.sh
# inference
sh 2-val.sh 

Main Results

Visualization Results

Instance predictions on LiDAR and corresponding images. Error maps.

Citation

If you find this project helpful in your research, please consider citing our paper:

@inproceedings{pan2025ial,
      title={How Do Images Align and Complement LiDAR? Towards a Harmonized Multi-modal 3D Panoptic Segmentation}, 
      author={Yining Pan and Qiongjie Cui and Xulei Yang and Na Zhao},
      booktitle = {Proceedings of the 42st International Conference on Machine Learning},
      year={2025},
      series = {ICML'25}
}

Acknowledgement

Many thanks to the following awesome open-source projects!

Closely Relevant Projects:

• LCPS
• P3Former

Other Awesome Projects:

• Grounding DINO
• Segment Anything Model
• nuScenes-devkit