This repository is the code for AD-VAT: An Asymmetric Dueling mechanism for learning Visual Active Tracking (ICLR 2019).
It contains the code for training/testing(Pytorch) and the 2D environments. The 3D environments are hosted in gym-unrealcv.
This repository requires:
- Python >= 3.6
- Pytorch >= 1.0
- Opencv >= 3.4
- Numpy == 1.14.0
- setproctitle, scikit-image, imageio, TensorboardX
See requirements.txt for more details.
To download the repository and install the requirements, you can run as:
git clone https://github.com/zfw1226/active_tracking_rl.git
cd active_tracking_rl
pip install -r requirements.txt
Note that you need install OpenCV, Pytorch, and the 2D/3D environments additionally.
We provide various 2D and 3D environments to validate the effectiveness of AD-VAT.
The 2D environment is a matrix map where obstacles are randomly placed.
The 2D experiment can run on a CPU-only machine.
In the 2D environments, you can evaluate and quantify the effectiveness of AD-VAT in a few minutes.
To install 2D environments(gym-track2d), you need run:
pip install -e envs/gym_track2d
The 3D environments are built on Unreal Engine(UE4), which could be flexibly customized to simulate
real-world active tracking scenarios.
To run the 3D environments, GPU is necessary.
To install 3D environments, please follow the instructions in gym-unrealcv.
You can try AD-VAT in 2D environments by running:
python main.py --shared-optimizer --workers 16 --split --train-mode -1 --env Track2D-BlockPartialPZR-v0
Note that you need adjust the number of --workers according to the number of your cpu cores.
It is important to limit number of worker processes to number of cpu cores available
as too many processes (e.g. more than one process per cpu core available) will actually be detrimental
in training speed and effectiveness.
--split means that it will save the tracker and target model separately for further evaluation.
Besides, you can also run the two baseline methods referred in the paper.
To train tracker with Ram target:
python main.py --shared-optimizer --workers 16 --split --train-mode 0 --env Track2D-BlockPartialRam-v0
To train tracker with Nav target:
python main.py --shared-optimizer --workers 16 --split --train-mode 0 --env Track2D-BlockPartialNav-v0
To train tracker and target under Naive dueling:
python main.py --shared-optimizer --workers 16 --split --network maze-lstm --entropy-target 0.01 --aux none --env Track2D-BlockPartialAdv-v0
You can evaluate the tracker by running:
python gym_eval.py --env {ENV_NAME} --network tat-maze-lstm --load-tracker {PATH_TO_YOUR_TRACKER}
The ENV_NAME we used to evaluate in the paper is:
Track2D-BlockPartialNav-v0(Block-Nav),Track2D-BlockPartialRam-v0(Block-Ram),Track2D-MazePartialNav-v0(Maze-Nav).Track2D-MazePartialRam-v0(Maze-Ram),
If you use the the default setting while training, the PATH_TO_YOUR_MODLE should be logs/{ENV_NAME}/{DATE}/tracker-best.dat
You can also evaluate the effectiveness of the tracker-aware target by running with different trackers, as:
python gym_eval.py --env Track2D-BlockPartialAdv-v0 --network tat-maze-lstm --load-tracker {PATH_TO_YOUR_TRACKER} --load-target {PATH_TO_YOUR_TARGET}
You can try AD-VAT in 3D environments by running:
python main.py --shared-optimizer --workers 6 --split --network tat-cnn-lstm --rnn-out 256 --entropy-target 0.05 --sleep-time 30 --env UnrealTrack-DuelingRoomPZR-DiscreteColor-v4 --env-base UnrealTrack-DuelingRoomNav-DiscreteColor-v4 --gray --rescale --lr 0.0001
To train the baselines, you only need reset --env to UnrealTrack-DuelingRoomRam-DiscreteColor-v4 or UnrealTrack-DuelingRoomNav-DiscreteColor-v4, and set --train-mode to 0 meanwhile.
You can evaluate the tracker by running:
python gym_eval.py --env {ENV_NAME} --network cnn-lstm --gray --rescale --rnn-out 256 --load-tracker {PATH_TO_YOUR_TRACKER}
The ENV_NAME we used to evaluate in the paper is:
UnrealTrack-DuelingRoomNav-DiscreteColor-v4(DR Room),UnrealTrack-UrbanCityNav-DiscreteColor-v1(Urban City),UnrealTrack-SnowForestNav-DiscreteColor-v1(Snow Village),UnrealTrack-GarageNav-DiscreteColor-v0(Parking Lot)
You could monitor the performance while training using tensorboard:
tensorboard --logdir {PATH_TO_LOGS}
If you use the the default setting while training, PATH_TO_LOGS should be logs/{ENV_NAME}/{DATE}
If you found AD-VAT useful, please consider citing:
@inproceedings{zhong2018advat,
title={{AD}-{VAT}: An Asymmetric Dueling mechanism for learning Visual Active Tracking},
author={Fangwei Zhong and Peng Sun and Wenhan Luo and Tingyun Yan and Yizhou Wang},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=HkgYmhR9KX},
}
@article{zhong2021advat,
author={Zhong, Fangwei and Sun, Peng and Luo, Wenhan and Yan, Tingyun and Wang, Yizhou},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
title={AD-VAT+: An Asymmetric Dueling Mechanism for Learning and Understanding Visual Active Tracking},
year={2021},
volume={43},
number={5},
pages={1467-1482},
doi={10.1109/TPAMI.2019.2952590}
}
If you have any suggestion/questions, get in touch at zfw@pku.edu.cn.