Recent advances in sequence modeling have resulted in powerful architectures for discovering patterns across different data types; however, heavyweight Transformers are excessive for signals like 3D skeleton gait. Motivated by this mismatch, we seek models that strike a balance between temporal expressiveness and computational efficiency. We introduce MAXGait, a lightweight network for soft-biometric inference that predicts age and gender from sequences of 3D joint coordinates. Built on Mamba, MAXGait utilizes selective state-space dynamics to efficiently capture long-range dependencies in joint trajectories. Moving away from the original design, MAXGait features a compact topology that maintains accuracy within tight resource constraints. To enhance representation, we develop Multi-Hierarchical Focusing Attention, which jointly models micro-motions and global gait cycles in the skeleton. Extensive experiments demonstrate that MAXGait is competitive with state-of-the-art baselines for age–gender classification on extended 3D skeleton sequences. Specifically, MAXGait achieves an overall accuracy and macro-F1 score of 72.77% and 72.5% on our self-curated Nano3Dgait dataset under a subject-overlapping protocol, 83.15% and 66.17% on OUMVLP-Mesh, and 88.62% and 84% on CASIA-B. It also offers over 20% efficiency gains and a smaller parameter footprint compared to a similarly compressed vanilla Mamba, while maintaining strong predictive performance. Yet, the significant disparity in sample sizes between groups still poses a serious challenge to unbiased performance, with the lowest per-group F1-score dropping to 33% under severely underrepresented conditions.
Chang et al. (Mon,) studied this question.