Home

Awesome

LoGoNet

This repo is the official implementation of: LoGoNet: Towards Accurate 3D Object Detection with Local-to-Global Cross-Modal Fusion.

Paper

Framework

image

News

Algorithm Modules

detection
ā”œā”€ā”€ al3d_det
ā”‚   ā”œā”€ā”€ datasets
ā”‚   ā”‚   ā”‚ā”€ā”€ DatasetTemplate: the basic class for constructing dataset
ā”‚   ā”‚   ā”‚ā”€ā”€ augmentor: different augmentation during training or inference
ā”‚   ā”‚   ā”‚ā”€ā”€ processor: processing points into voxel space
ā”‚   ā”‚   ā”‚ā”€ā”€ the specific dataset module
ā”‚   ā”œā”€ā”€ models: detection model related modules
|   |   ā”‚ā”€ā”€ fusion: point cloud and image fusion modules
ā”‚   ā”‚   ā”‚ā”€ā”€ image_modules: processing images
ā”‚   ā”‚   ā”‚ā”€ā”€ modules: point cloud detector
ā”‚   ā”‚   ā”‚ā”€ā”€ ops
ā”‚   ā”œā”€ā”€ utils: the exclusive utils used in detection module
ā”œā”€ā”€ tools
ā”‚   ā”œā”€ā”€ cfgs
ā”‚   ā”‚   ā”‚ā”€ā”€ det_dataset_cfgs
ā”‚   ā”‚   ā”‚ā”€ā”€ det_model_cfgs
ā”‚   ā”œā”€ā”€ train/test/visualize scripts  
ā”œā”€ā”€ data: the path of raw data of different dataset
ā”œā”€ā”€ output: the path of trained model
al3d_utils: the shared utils used in every algorithm modules
docs: the readme docs for LoGoNet

Running

šŸ’„ This project relies heavily on Ceph storage. Please refer to your file storage system to modify the file path.

Main results

Performances on Waymo with AP/APH (L2)

*We report average metrics across all results. We provide training / validation configurations, pretrained models for all models in the paper. To access these pretrained models, please send us an email with your name, institute, a screenshot of the the Waymo dataset registration confirmation mail, and your intended usage. Please note that Waymo open dataset is under strict non-commercial license so we are not allowed to share the model with you if it will used for any profit-oriented activities. However, we can provide the logs.

ModelmAPH_L2VEH_L2PED_L2CYC_L2Log
LoGoNet-1frame (val)71.3871.21/70.7175.49/69.9474.53/73.48log
LoGoNet-3frames (val)74.8674.60/74.1778.62/75.7975.44/74.61log
LoGoNet-5frames (val)75.5475.84/75.3878.97/76.3375.67/74.91log
LoGoNet-5frames (test)77.1079.69/79.3081.55/78.9173.89/73.10Record
LoGoNet_Ens (test)81.0282.17/81.7284.27/81.2880.93/80.06Record
LoGoNet_Ens_v2 (test)81.9682.75/82.3284.96/82.1082.36/81.46Record

Performances on KITTI with mAP

*We report average metrics across all results. We provide training / validation configurations, pretrained models for all models in the paper.

ModelCar@40Ped@40Cyc@40Log
LoGoNet (val)87.1364.4679.84log | weights
LoGoNet (test)85.8748.5773.61Record

Acknowledgement

We sincerely appreciate the following open-source projects for providing valuable and high-quality codes:

Reference

If you find our paper useful, please kindly cite us via:

@inproceedings{logonet,
  title={LoGoNet: Towards Accurate 3D Object Detection with Local-to-Global Cross-Modal Fusion},
  author={Xin Li and Tao Ma and Yuenan Hou and Botian Shi and Yuchen Yang and Youquan Liu and Xingjiao Wu and Qin Chen and Yikang Li and Yu Qiao and Liang He},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2023}
}

Contact