Dual-Branch Point-Supervised Action Detection Network with High Quality Pseudo-Labels

Point-supervised temporal action detection, which relies on sparse timestamp annotations, mitigates the annotation costs required by the fully supervised method. However, there is still a gap between point supervised methods and the fully supervised methods due to i) the sparsity of pseudo-labels and ii) insufficient exploitation of temporal information. We introduce a dual-branch architecture enhanced by high-quality pseudo-labels to narrow the gap. Specifically, this study proposes: 1) an attention-guided label propagation mechanism that supplements pseudo-labels by incorporating Transformer-generated attention map, 2) a Mamba-based pseudo-label generation mechanism which enhances local temporal pattern recognition to improve the reliability of pseudo-labels, and 3) a global completeness constraint to enforce temporal coherence and action integrity. Extensive experiments on three benchmark datasets validate the superiority of our approach, achieving state-of-the-art performance under point supervision settings.