Reproduce “MONET: Multiview Semi-supervised Keypoint Detection via Epipolar Divergence” 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
Yuan Yao, Yasamin Jafarian, Hyun Soo Park
Researchers at the University of Minnesota developed MONET, a semi-supervised framework that leverages a novel differentiable epipolar divergence loss to enforce geometric consistency directly on keypoint heatmaps. This approach enables accurate keypoint detection across diverse species using minimal labeled data, outperforming existing baselines in accuracy and precision.
Key insights
- Geometric consistency can be effectively enforced on keypoint heatmaps in an end-to-end differentiable manner by reparametrizing 2D distributions into 1D along rectified epipolar lines.
- The proposed epipolar divergence, a KL-divergence-inspired measure, provides a robust and computationally efficient way to quantify and minimize geometric inconsistencies across views.
- Integrating a small fraction of labeled data with a large amount of unlabeled multiview data, supervised by both epipolar divergence and 3D-derived pseudo-labels, dramatically reduces the need for manual annotation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.