4 November 2025 in Expert Systems with Applications
14 November 2025 in Knowledge-Based Systems
7 March 2026 in SSRN Electronic Journal
Hand keypoint detection in the wild is challenging due to the complexity of real-world conditions and the scarcity of labeled data. Although domain adaptation methods attempt to mitigate annotation costs by leveraging synthetic data, their dependence on controlled lab environments often hinders generalization to real-world scenarios. In this paper, we propose a novel semi-supervised framework named WildHand that leverages large-scale unlabeled data to simultaneously reduce annotation dependency and enhance generalization. First, we propose an attention-based dynamic filtering module that progressively integrates hard samples into training while reducing noise from unreliable pseudo labels through confidence-based weighting. Second, we introduce a reliability-guided co-teaching mechanism, where the more confident branch is selected as a temporary teacher with stop-gradient to provide a stable one-way supervision signal for the other branch, improving training stability and mitigating error reinforcement on unlabeled data. Moreover, we propose a hybrid data augmentation strategy named CutMax, and further align dense heatmap supervision with synthetic occlusions via an occlusion-masked training objective, preventing spurious gradients from occluded regions while increasing the diversity of training samples. Extensive experiments conducted on the COCO WholeBody Hand dataset demonstrate that WildHand consistently outperforms existing state-of-the-art methods. Furthermore, WildHand also demonstrates strong cross-domain generalization by reaching 59.1% PCK@0.05 on FreiHand and 83.9% PCK@0.05 on H3D, maintaining competitive performance without additional fine-tuning or adaptation.
2 September 2026 in Expert Systems with Applications