/VOTR

Voxel Transformer for 3D object detection

Primary LanguagePython

Voxel Transformer

This is a reproduced repo of Voxel Transformer for 3D object detection.

The code is mainly based on OpenPCDet.

Introduction

We provide code and training configurations of VoTr-SSD/TSD on the KITTI and Waymo Open dataset. Checkpoints will not be released.

Important Notes: VoTr generally requires quite a long time (more than 60 epochs on Waymo) to converge, and a large GPU memory (32Gb) is needed for reproduction. Please strictly follow the instructions and train with sufficient number of epochs. If you don't have a 32G GPU, you can decrease the attention SIZE parameters in yaml files, but this may possibly harm the performance.

Requirements

The codes are tested in the following environment:

  • Ubuntu 18.04
  • Python 3.6
  • PyTorch 1.5
  • CUDA 10.1
  • OpenPCDet v0.3.0
  • spconv v1.2.1

Installation

a. Clone this repository.

git clone https://github.com/PointsCoder/VOTR.git

b. Install the dependent libraries as follows:

  • Install the dependent python libraries:
pip install -r requirements.txt 
  • Install the SparseConv library, we use the implementation from [spconv].
    • If you use PyTorch 1.1, then make sure you install the spconv v1.0 with (commit 8da6f96) instead of the latest one.
    • If you use PyTorch 1.3+, then you need to install the spconv v1.2. As mentioned by the author of spconv, you need to use their docker if you use PyTorch 1.4+.

c. Compile CUDA operators by running the following command:

python setup.py develop

Training

All the models are trained with Tesla V100 GPUs (32G). The KITTI config of votr_ssd is for training with a single GPU. Other configs are for training with 8 GPUs. If you use different number of GPUs for training, it's necessary to change the respective training epochs to attain a decent performance.

The performance of VoTr is quite unstable on KITTI. If you cannnot reproduce the results, remember to run it multiple times.

  • models
# votr_ssd.yaml: single-stage votr backbone replacing the spconv backbone
# votr_tsd.yaml: two-stage votr with pv-head
  • training votr_ssd on kitti
CUDA_VISIBLE_DEVICES=0 python train.py --cfg_file cfgs/kitti_models/votr_ssd.yaml
  • training other models
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 sh scripts/dist_train.sh 8 --cfg_file cfgs/waymo_models/votr_tsd.yaml
  • testing
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 sh scripts/dist_test.sh 8 --cfg_file cfgs/waymo_models/votr_tsd.yaml --eval_all

Citation

If you find this project useful in your research, please consider cite:

@article{mao2021voxel,
  title={Voxel Transformer for 3D Object Detection},
  author={Mao, Jiageng and Xue, Yujing and Niu, Minzhe and others},
  journal={ICCV},
  year={2021}
}