Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Active Learning for GCN-based Action Recognition

Hichem Sahbi

TL;DR

This work tackles skeleton-based action recognition under scarce labels by coupling a probabilistic exemplar design within an active-learning loop to a label-efficient GCN. It introduces a display model that jointly optimizes representativeness, diversity, and uncertainty to select informative unlabeled samples, and a stable, invertible bidirectional GCN that enables robust mapping between ambient and latent spaces for exemplar design. A key theoretical contribution is the KM-Lipschitz stability analysis and practical regularization via orthonormal weight constraints, complemented by weight reparametrization to ensure numerical stability. Empirical results on SBU and FPHA datasets show that latent-space exemplar design within this framework yields clear gains at reduced labeling budgets, outperforming random, uncertainty-based, and core-set baselines. Overall, the approach offers a principled, data-efficient pathway for high-performance skeleton action recognition in annotation-constrained settings.

Abstract

Despite the notable success of graph convolutional networks (GCNs) in skeleton-based action recognition, their performance often depends on large volumes of labeled data, which are frequently scarce in practical settings. To address this limitation, we propose a novel label-efficient GCN model. Our work makes two primary contributions. First, we develop a novel acquisition function that employs an adversarial strategy to identify a compact set of informative exemplars for labeling. This selection process balances representativeness, diversity, and uncertainty. Second, we introduce bidirectional and stable GCN architectures. These enhanced networks facilitate a more effective mapping between the ambient and latent data spaces, enabling a better understanding of the learned exemplar distribution. Extensive evaluations on two challenging skeleton-based action recognition benchmarks reveal significant improvements achieved by our label-efficient GCNs compared to prior work.

Active Learning for GCN-based Action Recognition

TL;DR

This work tackles skeleton-based action recognition under scarce labels by coupling a probabilistic exemplar design within an active-learning loop to a label-efficient GCN. It introduces a display model that jointly optimizes representativeness, diversity, and uncertainty to select informative unlabeled samples, and a stable, invertible bidirectional GCN that enables robust mapping between ambient and latent spaces for exemplar design. A key theoretical contribution is the KM-Lipschitz stability analysis and practical regularization via orthonormal weight constraints, complemented by weight reparametrization to ensure numerical stability. Empirical results on SBU and FPHA datasets show that latent-space exemplar design within this framework yields clear gains at reduced labeling budgets, outperforming random, uncertainty-based, and core-set baselines. Overall, the approach offers a principled, data-efficient pathway for high-performance skeleton action recognition in annotation-constrained settings.

Abstract

Despite the notable success of graph convolutional networks (GCNs) in skeleton-based action recognition, their performance often depends on large volumes of labeled data, which are frequently scarce in practical settings. To address this limitation, we propose a novel label-efficient GCN model. Our work makes two primary contributions. First, we develop a novel acquisition function that employs an adversarial strategy to identify a compact set of informative exemplars for labeling. This selection process balances representativeness, diversity, and uncertainty. Second, we introduce bidirectional and stable GCN architectures. These enhanced networks facilitate a more effective mapping between the ambient and latent data spaces, enabling a better understanding of the learned exemplar distribution. Extensive evaluations on two challenging skeleton-based action recognition benchmarks reveal significant improvements achieved by our label-efficient GCNs compared to prior work.

Paper Structure

This paper contains 11 sections, 2 theorems, 7 equations, 6 tables.

Table of Contents

  1. Introduction
  2. Display Model
  3. Display model design
  4. Learning Model
  5. Graph convnets at a glance
  6. Proposed stable bidirectional GCNs
  7. Weight reparametrization
  8. Experiments
  9. Display model: comparison & ablation
  10. Regularization and weight reparametrization
  11. Conclusion

Key Result

Proposition 1

The optimality conditions of Eq. of yield the following iterative update of the solution as a fixed point of where $\hat{\mathbf{\mu}}^{(\tau+1)}$ and $\hat{\mathbf{V}}^{(\tau+1)}$ are given by with the similarity matrix $\mathbf{S}$ between $\mathbf{V}$ and $\mathbf{H}$ (where $\mathbf{V}^{(\tau)}$ is denoted as $\mathbf{V}$ for brevity) defined as here $\mathbf{1}_N$ is a vector of $N$ ones,

Theorems & Definitions (3)

  • Proposition 1
  • Definition 1: Stability
  • Proposition 2