Slow-fast frequency decomposition for forecasting human hand articulation

Forecasting 3D hand articulation is inherently ambiguous. The high dimensionality of hand kinematics and the coexistence of slow global motion with rapid, asynchronous finger movements imply that identical observation windows may correspond to multiple plausible futures. In this regime, greater architectural complexity alone is insufficient and often ineffective; meaningful progress requires inductive biases that reflect the intrinsic temporal structure of hand motion. We propose SpecSFNet , the Spectral Slow-Fast Network, a lightweight trajectory-space model that introduces an explicit slow-fast decomposition in the discrete cosine transform (DCT) domain to cope with the intrinsic temporal heterogeneity of hand motion. By enforcing a frequency-aware inductive bias that disentangles low-frequency global dynamics from high-frequency finger articulation and recombines them adaptively, the model enables stable and accurate trajectory reconstruction where uniform temporal modeling proves insufficient. Experiments on four public benchmarks (AssemblyHands, H2O, FPHA, BigHand2.2M) show that SpecSFNet consistently improves the MPJPE metric over strong full-body forecasting methods adapted to hands, including MLP, transformer, graph-based, and equivariant models, while remaining computationally efficient for real-time inference. Ablation studies indicate that frequency-aware temporal factorization, rather than increased model complexity, is the primary driver of improved accuracy and temporal coherence. Code is available at https://anonymous.4open.science/r/4D_hands-5975/README.md .