Sequential Joint Dependency Aware Human Pose Estimation with State Space Model

Authors: Hanxi Yin, Shaodi You, Jungong Han, Zhixiang Chen

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments are conducted on two datasets to validate the effectiveness of our proposed SSM module, and the results demonstrate that our pose estimator can deliver impressive performance.
Researcher Affiliation Academia 1University of Amsterdam 2University of Sheffield EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Joint-Dependent SSM Layer Input: Fx: (B, L, Ds) Output: Fy: (B, L, Ds) 1: A: (Ds, (N) Parameter Represents structured N N matrix 2: B: (L, N) Parameter 3: C: (L, N) Parameter 4: : (L, Ds) Softplus(Parameter) 5: A, B: (L, Ds, N) discretize( , A, B) 6: h: (B, Ds, N) initialize( B, Fx) with Eq. 2 7: Fy SSM( A, B, C)(Fx, h) with Eqs. 3 and 4 Joint-dependent 8: return Fy
Open Source Code Yes Code https://github.com/yinhanxi/Pose SSM
Open Datasets Yes Our method is evaluated on Human3.6M (Ionescu et al. 2014) and MPI-INF-3DHP (Mehta et al. 2017a).
Dataset Splits Yes Human3.6M. Following previous works (Gong, Zhang, and Feng 2021; Wandt and Rosenhahn 2019; Martinez et al. 2017; Zeng et al. 2020; Chen et al. 2018), we utilize subjects 1, 5, 6, 7 and 8 for training, and subjects 9 and 11 for evaluation.
Hardware Specification Yes Our method is implemented on a single NVIDIA Ge Force RTX 4090 GPU.
Software Dependencies No The paper mentions that the method is implemented on a single NVIDIA Ge Force RTX 4090 GPU but does not specify software dependencies like programming language versions or library versions (e.g., PyTorch version, CUDA version).
Experiment Setup Yes We train the model 30 epochs with a batch size of 512. The learning rate is initialized at 0.0005, decayed by 0.95 per epoch and halved every 5 epochs. Horizontal flip is applied as data augmentation in training. Experimentally, we set D, K and N as 160, 2 and 4 with the optimal MPJPE on Human3.6M by searching each parameter independently. We search D from 96 to 240, with a step size of 16, K from {2,4,8} and N from {2,4,8,16}. We initialize SSM parameters A and according to Mamba (Gu and Dao 2023), and randomly initialize B and C.