Awesome

Pose-Assisted Multi-Camera Collaboration for Active Object Tracking

This repository is the python implementation of Pose-Assisted Multi-Camera Collaboration for Active Object Tracking (AAAI 2020).

It contains the code for training/testing(Pytorch). The 3D environments are hosted in gym-unrealcv.

Dependencies

The repository requires:

• Linux (Ubuntu 16)
• Python >= 3.6
• Pytorch >= 1.0
• gym-unrealcv >= 1.0
• OpenCV >= 3.4
• Numpy == 1.14.0
• setproctitle, scikit-image, imageio, TensorboardX, Matplotlib

Prepare 3D Environments

Install gym-unrealcv:

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

Load environment binaries:

python load_env.py -e Textures
python load_env.py -e MCRoom
python load_env.py -e UrbanTree
python load_env.py -e Garden

Installation

To download the repository and install the requirements, you can run as:

git clone https://github.com/LilJing/pose-assisted-collaboration.git
cd pose-assisted-collaboration
pip install -r requirements.txt

Note that you need install OpenCV, Pytorch, and the 3D environments additionally.

Training

Train the vision-based controller

Use the following command:

python main.py --rescale --shared-optimizer --env UnrealMCRoom-DiscreteColorGoal-v5 --workers 6

Train the pose-based controller

cd ./pose
python main.py --env PoseEnv-v0 --shared-optimizer --workers 12

The best parameters of the network will be saved in corresponding logs dir.

Evaluation

There are two environments for evaluation, Garden and Urban City.

We provide the pre-trained model in .models/. The trained vision-based controller model is Vision-model-best.dat and the pose-based controller model is Pose-model-best.dat.

Run our model on Garden:

python evaluate.py --rescale --load-vision-model  ./models/Vision-model-best.dat --load-pose-model ./models/Pose-model-best.dat  --env UnrealGarden-DiscreteColorGoal-v1 --num-episodes 100 --test-type modelgate --render

Run our model on Urban City:

python evaluate.py --rescale --load-vision-model  ./models/Vision-model-best.dat --load-pose-model ./models/Pose-model-best.dat  --env UnrealUrbanTree-DiscreteColorGoal-v1 --num-episodes 100 --test-type modelgate --render

Demo Videos

To see demo videos, please refer to YouTube.

Citation

If you found this work useful, please consider citing:

@inproceedings{li2020pose,
  title={Pose-Assisted Multi-Camera Collaboration for Active Object Tracking},
  author={Jing Li*, Jing Xu*, Fangwei Zhong*, Xiangyu Kong, Yu Qiao,  Yizhou Wang},
  booktitle={The Thirty-Fourth AAAI Conference on Artificial Intelligence},
  year={2020}
}