Exploiting multi-transformer encoder with multiple-hypothesis aggregation via diffusion model for 3D human pose estimation

Abstract

The transformer architecture has consistently achieved cutting-edge performance in the task of 2D to 3D lifting human pose estimation. Despite advances in transformer-based methods they still suffer from issues related to sequential data processing, addressing depth ambiguity, and effective handling of sensitive noisy data. As a result, transformer encoders encounter difficulties in precisely estimating human positions. To solve this problem, a novel multi-transformer encoder with a multiple-hypothesis aggregation (MHAFormer) module is proposed in this study. To do this, a diffusion module is first introduced that generates multiple 3D pose hypotheses and gradually distributes Gaussian noise to ground truth 3D poses. Subsequently, the denoiser is employed within the diffusion module to restore the feasible 3D poses by leveraging the information from the 2D keypoints. Moreover, we propose the multiple-hypothesis aggregation with a join-level reprojection (MHAJR) approach that redesigns the 3D hypotheses into the 2D position and selects the optimal hypothesis by considering reprojection errors. In particular, the multiple-hypothesis aggregation approach tackles depth ambiguity and sequential data processing by considering various possible poses and combining their strengths for a more accurate final estimation. Next, we present the improved spatial-temporal transformers encoder that can help to improve the accuracy and reduce the ambiguity of 3D pose estimation by explicitly modeling the spatial and temporal relationships between different body joints. Specifically, the temporal-transformer encoder introduces the temporal constriction & proliferation (TCP) attention mechanism and the feature aggregation refinement module (FAR) into the refined temporal constriction & proliferation (RTCP) transformer, which enhances intra-block temporal modeling and further refines inter-block feature interaction. Finally, the superiority of the proposed approach is demonstrated through comparison with existing methods using the Human3.6M and MPI-INF-3DHP benchmark datasets.