Reproduce “Particle Video Revisited: Tracking Through Occlusions Using Point Trajectories” 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
Adam W. Harley, Zhaoyuan Fang, Katerina Fragkiadaki
Researchers at Carnegie Mellon University introduce Persistent Independent Particles (PIPs), a method that extends the classic particle video concept to robustly track pixel trajectories through occlusions. PIPs achieves superior performance on synthetic and real-world datasets, particularly for occluded targets, by learning strong multi-frame temporal priors.
Key insights
- Dedicate significant model capacity to multi-frame temporal reasoning for individual particles, trading off spatial smoothness for temporal robustness.
- An iterative deep temporal network (MLP-Mixer) can learn powerful priors to re-acquire targets after occlusions and infer amodal trajectories.
- Synthetic data augmentation, such as the created FlyingThings++, is crucial for supervising models to learn amodal tracking through occlusions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.