Reproduce “Deep kernel video approximation for unsupervised action segmentation” 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
Silvia L. Pintea, Jouke Dijkstra
This research introduces a method for per-video unsupervised action segmentation by learning a compact approximation of a video's frame distribution using Maximum Mean Discrepancy (MMD) with Neural Tangent Kernels (NTKs). The approach achieved competitive and often superior F1 scores compared to state-of-the-art baselines, particularly when the true number of action segments was unknown.
Key insights
- The combination of a Gaussian kernel with an NTK (Gauss x NTK) provides superior segmentation accuracy, effectively leveraging both global characteristic properties and local descriptive power.
- Learning a video approximation by minimizing MMD in a deep kernel space enables robust action segmentation, even when the true number of segments is unknown, by adapting to the underlying data distribution.
- Maximum Mean Discrepancy (MMD) offers a computationally efficient and theoretically sound alternative to Optimal Transport for geometry-preserving distribution comparisons in unsupervised video analysis.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.