Reproduce “Exploiting Temporal Contexts with Strided Transformer for 3D Human Pose Estimation” on OpenResearch

About this paper

Wenhao Li, Hong Liu, Runwei Ding, Mengyuan Liu, Pichao Wang, Wenming Yang

Researchers developed the Strided Transformer, an architecture for 3D human pose estimation that efficiently leverages temporal contexts from video sequences using a Strided Transformer Encoder and a full-to-single supervision scheme. This method achieved state-of-the-art results on the Human3.6M dataset, with a 43.7 mm MPJPE, and improved temporal smoothness by reducing MPJVE by 15.4% compared to a variant without full sequence supervision.

Key insights

• Video sequences contain redundant 2D pose information, which can be efficiently exploited through hierarchical compression rather than being processed individually at full length.
• Integrating strided 1D convolutions within the Transformer's feed-forward network in the Strided Transformer Encoder can effectively reduce computational complexity and capture local features, complementing global self-attention.
• A dual supervision approach, optimizing for both full-sequence temporal consistency and single-target-frame accuracy, is crucial for producing both smooth and precise 3D human pose estimations.

Summary by alphaXiv.