Reproduce “Learning Target Candidate Association to Keep Track of What Not to Track” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Christoph Mayer, Martin Danelljan, Danda Pani Paudel, Luc Van Gool
Researchers from ETH Zurich developed KeepTrack, a visual object tracking method that enhances robustness by actively associating and managing both target and distractor identities. It achieved new state-of-the-art performance on six challenging benchmarks, including a +4.0% AUC gain on LaSOT over its baseline.
Key insights
- Actively tracking and understanding the identities and movements of non-target (distractor) objects provides invaluable context, enabling more confident target identification.
- A learned association network can effectively propagate identities of all target candidates (including distractors) over time, even with incomplete ground-truth annotations through a hybrid supervision strategy.
- A dynamic memory management strategy for online learning, which incorporates confidence derived from candidate association, enhances tracker adaptation and prevents model drift.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.