HM-ViT

[ICCV 2023] HM-ViT: Hetero-modal Vehicle-to-Vehicle Cooperative perception with vision transformer

This is the official implementation of ICCV 2023 paper "HM-ViT: Hetero-modal Vehicle-to-Vehicle Cooperative perception with vision transformer". Hao Xiang, Runsheng Xu, Jiaqi Ma

Installation

Please refer to Bevformer to install dependancies for mmdetection, mmdet3d, and mmcv.

# Clone repo
git clone https://github.com/XHwind/HM-ViT

cd opencood

# Setup conda environment
conda create -y --name v2xvit python=3.7

conda activate v2xvit
# pytorch >= 1.8.1, newest version can work well
conda install -y pytorch torchvision cudatoolkit=11.3 -c pytorch
# spconv 2.0 install, choose the correct cuda version for you
pip install spconv-cu113

# Install dependencies
pip install -r requirements.txt
# Install bbx nms calculation cuda version
python v2xvit/utils/setup.py build_ext --inplace

# install opencood into the environment
python setup.py develop

Data

Please follow the instruction in OPV2V to download the data.

Quick start

Please change the following parameters in the config to configure different hetero modes and ego types:

camera_to_lidar_ratio: 0
ego_mode: 'lidar'

camera_to_lidar_ratio controls the ratio of collaborators's camera/LiDAR ratio. camera_to_lidar_ratio=0 corresponds to pure lidar collaborators while camera_to_lidar_ratio=1 is pure camera collaborators.

ego_mode specifies the ego agent's modality. ego_mode=lidar is lidar ego agent while ego_mode=camera is camera ego agent, ego_mode=mixed will have the equal probability of each modality.

To train with multiple gpus:

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4  --use_env opencood/tools/train_camera.py --hypes_yaml opencood/hypes_yaml/opcamera/cvt.yaml --model_dir opencood/logs/cvt_att_fuse

To inference:

python -m opencood.tools.inference_camera --model_dir path/to/log \
    --fusion_method \
    nofusion \
    --ego_mode \
    camera \
    --camera_to_lidar_ratio \
    1 \
    --ego_mode \
    mixed \