Sequence Modeling and Generative Model Driven Non-Rigid 3D Reconstruction

Non-rigid 3D reconstruction is a critical problem in the computer vision community. Although researchers have made significant progress in recent years, two key challenges remain: (1) ensuring temporal consistency in the reconstructed results, and (2) achieving accurate reconstruction under uncertainty in the input data. To address these challenges, we propose a novel framework that integrates a Temporal Convolutional Network (TCN) for modeling sequential dependencies and ensuring temporal consistency, alongside a diffusion-based module that generates pseudo-3D structures to provide teacher supervision. This diffusion-based supervision enhances spatial accuracy, particularly under weakly supervised conditions. Extensive experiments on the Human3.6M dataset show that our approach achieves superior reconstruction performance compared to existing baselines.