Reproduce “PointOdyssey: A Large-Scale Synthetic Dataset for Long-Term Point 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
Yang Zheng, Adam W. Harley, Bokui Shen, Gordon Wetzstein, Leonidas J. Guibas
Stanford University researchers introduce PointOdyssey, a large-scale synthetic dataset for long-term point tracking, alongside PIPs++, an improved tracking algorithm. Training on PointOdyssey enhances tracking performance on real-world datasets, reducing TAP-Vid-DAVIS MTE from 6.30 to 5.14, and significantly boosts inference speed from 4 FPS to 55 FPS.
Key insights
- Naturalistic synthetic data, incorporating real-world motion capture and detailed scene modeling, effectively bridges the sim-to-real gap for long-term point tracking.
- Leveraging expanded temporal context through convolutional architectures and multi-template appearance matching improves tracking robustness and efficiency over long video sequences.
- The PointOdyssey dataset's design, including physical interactions and ego-centric views, provides a challenging benchmark that pushes the boundaries of current tracking capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.