Baseline methods for object tracking simply compute IoU over consecutive frames to track objects detected by bounding boxes.
Here I tried to use the feature map associated with each bounding boxes instead, as a proof of concept. I took the time to document each parts of the code, which coupled with the simplicity of the task might make it a good introductory example to deep learning.
Result example from a random youtube video
This will download ~200Mb of example data, ~100Mb of example results and a bit of code. You could run the script in a bare ubuntu docker image, or in your virtual environment of choice.
./setup.shOpen the jupyter notebook (how to install jupyter) Tracking_using_feature_maps.ipynb for details about the method.
- Dockerize
- Add one command script to track objects