Skip to content

Latest commit

 

History

History
87 lines (59 loc) · 3.28 KB

README.md

File metadata and controls

87 lines (59 loc) · 3.28 KB
Raw

Active tracking toolkit

This is a simple framework for evaluating tracking algorithms on 3D environments. Our 3D environments is built in Unreal Engine 4, and we utilize gym_unrealcv as a python wrapper.

Usage

Install gym_unrealcv

First, install our modified gym_unrealcv module.

cd envs/gym-unrealcv
pip install -e .

Prepare environments

We provide 2 environments available:

  • GeometryTrack (UnrealTrack-GeometryTrackRam-DiscreteColor-v0)
  • ForestTrack (UnrealTrack-ForestTrack-DiscreteColor-v0)

You can download the environments here, unzip and put them in ./envs/gym_unrealcv/envs/UnrealEnv. Then, modify the env_bin to the binary excutable file and img_dir to the textures directory in the gym_unrealcv environment setting file.

Taking 'GeometryTrack' for example, the setting file is ./envs/gym-unrealcv/gym_unrealcv/envs/setting/track/GeometryTrack.json, modified terms should be like:

"env_bin": "/home/username/activate_tracking_toolkit/envs/gym-unrealcv/gym_unrealcv/envs/UnrealEnv/GeometryTrack2/GeometryTrack2/Binaries/Linux/GeometryTrack2",
"imgs_dir": "/home/username/activate_tracking_toolkit/envs/gym-unrealcv/gym_unrealcv/envs/UnrealEnv/textures",

According to gym_unrealcv, relative path can be used, but somehow, absolute path is necessary here.

If you want to customize your own environments, please refer to tutorial of gym_unrealcv.

Evaluate trackers

You can run test.py to test you tracker and use --env to specify the test environment. An example is as follows:

python test.py --env UnrealTrack-GeometryTrackRam-DiscreteColor-v0 --model_path [your model path]

For other options available, you can refer to the source code.

For trackers already integrated in our toolkit, you can get the pretrained weights from reference links.

Add new trackers

If you want to test your own passive tracking method, you need to inherit the TrackerBase class defined in tracker.py. We follow a common setting that you need to override a init() function to get the tracking target as input and initialize your tracker, and a track() function to perform tracking.

Then, you can put your source code in the trackers folder and modify __init__.py, so that when testing, you can import your tracker class like:

from trackers import TrackerSiamFC

You can refer to the SiamFC implementation as an example.

Reference

UE4 environment wrapper: https://github.com/zfw1226/gym-unrealcv

SiamFC: https://github.com/huanglianghua/siamfc-pytorch

Transformer tracking(TransT): https://github.com/chenxin-dlut/TransT