Feature Aware Re-weighting (FAR) in Bird’s Eye View for LiDAR-based 3D object detection in autonomous driving applications
This is the official implementation of "Feature Aware Re-weighting (FAR) in Bird’s Eye View for LiDAR-based 3D object detection in autonomous driving applications" paper, that you can download here. This project is built on OpenPCDet.
All the codes are tested in the following environment:
- Linux (tested on Ubuntu 18.04)
- Python 3.8.8
- PyTorch 1.10
- CUDA 11.1
- Install the spconv library from spconv.
- Install pytorch 1.10
conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.1 -c pytorch -c conda-forge - Install requirements
pip install -r requirements.txt - Install pcdet library
python setup.py develop
Selected supported methods are shown in the below table. The results are the 3D detection performance of moderate difficulty on the val set of KITTI dataset. All models are trained on a single RTX 3090 GPU.
| Car@R11 | Pedestrian@R11 | Cyclist@R11 | weights | |
|---|---|---|---|---|
| PointPillar_FAR | 76.87 | 52.05 | 63.63 | model |
| CenterPoint_FAR | 76.73 | 50.72 | 65.10 | model |
| SECOND_FAR | 78.30 | 53.92 | 67.27 | model |
| PV-RCNN_FAR | 83.38 | 60.43 | 72.47 | model |
| Voxel R-CNN_FAR (all classes) | 83.89 | 60.76 | 72.18 | model |
By default, all models are trained with a single frame of 20% data (~32k frames) of all the training samples on a single RTX 3090 GPU, and the results of each cell here are mAP calculated by the official Waymo evaluation metrics on the whole validation set.
| Performance@(train with 20% Data) | Vec_L1 | Vec_L2 | Ped_L1 | Ped_L2 | Cyc_L1 | Cyc_L2 |
|---|---|---|---|---|---|---|
| PointPillar_FAR | 71.30 | 63.02 | 67.15 | 58.90 | 58.26 | 56.06 |
| CenterPoint_FAR-Dynamic-Pillar | 71.49 | 63.24 | 74.30 | 66.20 | 66.63 | 64.13 |
| SECOND_FAR | 71.13 | 62.86 | 65.78 | 57.83 | 59.18 | 56.99 |
| PV-RCNN_FAR (AnchorHead) | 75.17 | 66.59 | 72.19 | 63.17 | 67.27 | 64.76 |
| Voxel R-CNN_FAR (CenterHead)-Dynamic-Voxel | 76.18 | 67.76 | 77.95 | 69.28 | 71.15 | 68.53 |
All models are trained on a single RTX 3090 GPU and are available for download.
| mATE | mASE | mAOE | mAVE | mAAE | mAP | NDS | download | |
|---|---|---|---|---|---|---|---|---|
| PointPillar_FAR-MultiHead | 33.88 | 25.99 | 31.73 | 28.57 | 20.24 | 45.38 | 58.65 | model |
| SECOND_FAR-MultiHead (CBGS) | 31.64 | 25.56 | 27.14 | 25.48 | 19.84 | 51.31 | 62.69 | model |
| CenterPoint_FAR-PointPillar | 30.94 | 25.81 | 41.52 | 23.55 | 19.36 | 52.05 | 61.91 | model |
If you find this work useful in your research, please consider cite:
@article{ZAMANAKOS2024104664,
title = {Feature Aware Re-weighting (FAR) in Bird’s Eye View for LiDAR-based 3D object detection in autonomous driving applications},
journal = {Robotics and Autonomous Systems},
pages = {104664},
year = {2024},
issn = {0921-8890},
doi = {https://doi.org/10.1016/j.robot.2024.104664},
url = {https://www.sciencedirect.com/science/article/pii/S0921889024000472},
author = {Georgios Zamanakos and Lazaros Tsochatzidis and Angelos Amanatiadis and Ioannis Pratikakis}}