Improved version of OpenLabeling. Its two main contributions are:
- The integration of the state-of-the-art real-time class-agnostic object tracker SiamMask, which handles particularly well severe changes in appearance and bounding box aspect ratio.
- Enables to skip the tracking loop by pressing the key
P(or the spacebar).
Those two features combined enable a significant time-saving in the tedious process of labeling sequences.
Works both on CPU (3-5fps) and GPU (20-30fps).
Manually provided bounding box of a vehicle at frame 0.
Bounding box of the tracked vehicle after 230 frames. Notice how SiamMask has automatically modified the aspect ratio of the vehicle after it turned.
- To install all the dependencies, simply type:
pip install -r requirements.txt
- Download the SiamMask pretrained weights:
cd siamMask/pretrained
wget -q http://www.robots.ox.ac.uk/~qwang/SiamMask_VOT.pth
wget -q http://www.robots.ox.ac.uk/~qwang/SiamMask_DAVIS.pth
- Run OpenLabeling's GUI (from the root folder):
python -m main.main
By default, SiamMask will be used. To use another tracker, provide it as command line argument, like so:
python -m main.main --tracker_type KCF