Modulated Individually Connected Layers for 3D Human Pose Estimation
Hao Wang, Yahui Gao, Ruijun Liu, Xiaochuan Wang · 2024
The regression head is a crucial component in 3D Human Pose Estimation for enhancing model performance. Most existing methods utilize a shared linear transformation matrix to estimate different joints, neglecting the matching information between joints and their corresponding vectors in feature maps. By incorporating insights from Individually Connected Layers and Modulated Graph Convolutional Networks, we propose Modulated Individually Connected Layers. This novel approach learns a distinct modulation vector for each joint by modulating the shared linear transformation matrix. Empirical experiments demonstrate that this modulation technique improves the performance of state-of-the-art (SOTA) models by approximately \(0.3\%\) to \(2.0\%\) on the Human3.6M dataset and \(1.0\%\) to \(12.1\%\) on the MPI-INF-3DHP dataset without hyperparameter fine-tuning. Further generalization and ablation studies verify the effectiveness of our module. Moreover, the proposed Modulated Individually Connected Layers provide a lightweight, easily integrable design that seamlessly enhances the performance of deterministic methods with minimal computational overhead.