DimHand: A Large-Scale Benchmark for
Hand Pose Estimation under
Dynamic Low-Light Conditions

RGB-based 3D hand pose estimation (HPE) in dynamic real-world environments is a foundational technology for human-computer interaction. Although current methods achieve remarkable performance under controlled conditions, they remain sensitive to extremely low-light conditions. Specifically, it involves low signal-to-noise ratios and missing visual cues, making it difficult to acquire precise 3D motion annotations from RGB. As a compromise, existing low-light datasets capture static poses, failing to reflect real-world complex dynamic motions. To this end, we develop a multi-modal, multi-view capture system with adjustable illumination and a high-fidelity illumination-invariance automated annotation pipeline, decoupling 3D annotation from adverse lighting, preserving photometric realism and illumination variations. Based on this pipeline, we propose DimHand, the first large-scale 3D hand pose dataset capturing continuous motion under dynamic low-light conditions, comprising 745,280 frames from 15 subjects performing 15 distinct sequential actions across 5 viewpoints with illumination ranging from 0.1 to 10 lux. However, the coupling between severe visual degradation and the intrinsic complexity of hand structure induces spatial and semantic ambiguity. To address these challenges, we propose DimNet, an end-to-end framework that leverages 3D structural priors via a Kinematic Prior Codebook (KPC) for structural regularization and a Spatial-Semantic Projection (SSP) for local semantic refinement. Extensive experiments demonstrate that DimNet outperforms state-of-the-art methods, particularly in generalizing to unseen dynamic hand motions, thereby establishing a strong low-light baseline for 3D HPE.