ADNet implementation in Python using PyTorch and PyTracking.
The tracker is trained using supervised learning and a modified REINFORCE policy gradient algorithm.
We show that using a curriculum speeds up reinforcement learning. The curriculum is built from synthetic sequences gradually increasing in difficulty.
This repository is part of my undergraduate thesis.
Tracking on a synthetic sequence, without fine tuning, trained using only reinforcement learning.
The tracker is incorporated into the (modified) PyTracking library.
- Tracker evaluation and training source code files are found in pytracking/pytracking/tracker/adnet/.
- Tracker parameter files are found in pytracking/pytracking/parameter/adnet/.
- Included is a simple synthetic sequence generator.
For setup instructions and a training demonstration see the example Jupyter notebook.
