Reproduce “Efficient Motion Prompt Learning for Robust Visual Tracking” 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
Jie Zhao, Xin Chen, Yongsheng Yuan, Michael Felsberg, Dong Wang, Huchuan Lu
"Motion Prompt-based Tracking" (MPT) augments visual object trackers by integrating explicit long-term motion trajectories as prompts. This plug-and-play module enhances tracking robustness, achieving a 10.2% Expected Average Overlap (EAO) gain on VOT2018, while drastically reducing training time to 5 hours on 6GB GPU memory compared to 38 hours for baseline models.
Key insights
- Explicit long-term motion trajectories can be effectively integrated as prompts to enhance visual tracking robustness, circumventing the need for computationally heavy sequential training.
- A modular design, featuring a specialized motion encoder and a Transformer-based fusion decoder with an adaptive weighting mechanism, allows for flexible and efficient integration of temporal and visual information.
- By freezing baseline tracker parameters and training only the new MPT components, the proposed prompt learning strategy achieves superior or comparable performance with substantially reduced training time and memory footprint.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.