Skip to content
Book Open access

Active Learning for Human Pose Estimation via Submodular Optimization on Latent Hyper-Ellipsoids

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 17 references

Abstract

While deep neural networks have significantly advanced 2D Human Pose Estimation, their performance relies heavily on massive datasets with costly, dense skeletal annotations. In practical scenarios, while raw data is abundant, the high costs of manual annotation necessitate budget-constrained approaches. Active learning mitigates this annotation bottleneck, but traditional heuristics often struggle with structural overconfidence and batch redundancy when dealing with complex, articulated poses. To address this, we propose a novel, density-aware active learning framework designed to robustly explore unannotated pose distributions. By leveraging a variational auto-encoder, we map unlabeled images into a probabilistic latent space, conceptualizing instances as hyper-ellipsoids. This representation explicitly captures both local data density and epistemic uncertainty, enabling the computation of a directed Mahalanobis divergence. We then ensure diverse and highly informative batch selection by optimizing a submodular facility-location function, weighted by a ''likely-miss'' prediction score. Experimental results on standard Human Pose Estimation benchmarks demonstrate that our geometric approach significantly accelerates convergence, outperforming current state-of-the-art active learning methods with substantially fewer labeled samples.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.