Reproduce “One Graph to Track Them All: Dynamic GNNs for Single- and Multi-View 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
Martin Engilberge, Ivan Vrkic, Friedrich Wilke Grosche, Julien Pilet, Engin Turetken, Pascal Fua
A unified, fully differentiable model named Unified Message Passing Network (UMPN) is introduced for online multi-people tracking in both single- and multi-view settings. The model incorporates scene structure for improved occlusion handling, achieving 93.9% MOTA on WILDTRACK and outperforming baselines on the new SCOUT dataset while processing at 23 FPS.
Key insights
- A dynamic graph representation and message passing can effectively unify spatio-temporal, contextual, and multi-view information for robust online tracking across diverse camera setups.
- Directly integrating 3D scene structure and visibility information through specialized graph nodes and edges significantly enhances tracking robustness against prolonged environmental occlusions.
- A "delayed backpropagation" training strategy is critical for enabling GNNs to learn long-term dependencies in dynamic graphs, which is essential for effective online inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.