Empowering Embodied Visual Tracking with Visual Foundation Models and Offline RL

FangWei Zhong, Kui Wu, Hai Ci, Chu-ran Wang , Hao Chen

Peking University, BeiHang University, National University of Singapore and The Hong Kong Polytechnic University.

ECCV 2024

[arXiV] [Project Page]

Installation

Our model rely on the DEVA as the vision foundation model and Gym-Unrealcv as the evaluation environment, which requires to install three additional packages: Grounded-Segment-Anything, DEVA and Gym-Unrealcv. Note that we modified the original DEVA to adapt to our task, we provide the modified version in the repository. Prerequisite:

• Python 3.9
• PyTorch 2.0.1+ and corresponding torchvision
• gym_unrealcv(https://github.com/zfw1226/gym-unrealcv)
• Grounded-Segment-Anything (https://github.com/hkchengrex/Grounded-Segment-Anything)
• DEVA (https://github.com/hkchengrex/Tracking-Anything-with-DEVA)

Clone our repository:

git clone https://github.com/wukui-muc/Offline_RL_Active_Tracking.git

Install Grounded-Segment-Anything:

cd Offline_RL_Active_Tracking
git clone https://github.com/hkchengrex/Grounded-Segment-Anything

export AM_I_DOCKER=False
export BUILD_WITH_CUDA=True
export CUDA_HOME=/path/to/cuda/
# /path/to/cuda/ is the path to the cuda installation directory, e.g., /usr/local/cuda
# if you install the cuda in conda, it should be {path_to_conda}/env/{conda_env_name}/lib/, e.g., ~/anaconda3/env/offline_evt/lib/

cd Grounded-Segment-Anything
python -m pip install -e segment_anything
python -m pip install -e GroundingDINO
pip install --upgrade diffusers[torch]

Install DEVA:
Directly install the modified DEVA in the repository
(If you encounter the File "setup.py" not found error, upgrade your pip with pip install --upgrade pip)

cd ../Tracking-Anything-with-DEVA # go to the DEVA directory
pip install -e .
bash scripts/download_models.sh #download the pretrained models

Install Gym-Unrealcv:

cd ..
git clone https://github.com/zfw1226/gym-unrealcv.git
cd gym-unrealcv
pip install -e .

Before running the environments, you need to prepare unreal binaries. You can load them from clouds by running load_env.py

python load_env.py -e {ENV_NAME}

# To run the demo evaluation script, you need to load the UrbanCityMulti environment and textures by running:
python load_env.py -e UrbanCityMulti
python load_env.py -e Textures
sudo chmod -R 777 ./   #solve the permission problem

Quick Start

Training

python train_offline --buffer_path {Data-Path}

Evaluation

python Eval_tracking_agent.py --env UnrealTrackGeneral-UrbanCity-ContinuousColor-v0 --chunk_size 1 --amp --min_mid_term_frames 5 --max_mid_term_frames 10 --detection_every 20 --prompt person.obstacles 

Citation

@inproceedings{zhong2024empowering,
  title={Empowering Embodied Visual Tracking with Visual Foundation Models and Offline RL},
  author={Zhong, Fangwei and Wu, Kui and Ci, Hai and Wang, Churan and Chen, Hao},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2024}
}

References

Thanks for the previous works that we build upon:
DEVA: https://github.com/hkchengrex/Tracking-Anything-with-DEVA
Grounded Segment Anything: https://github.com/IDEA-Research/Grounded-Segment-Anything
Segment Anything: https://github.com/facebookresearch/segment-anything
XMem: https://github.com/hkchengrex/XMem
Title card generated with OpenPano: https://github.com/ppwwyyxx/OpenPano