A 3D Human Pose Estimation Method Based on Adaptive Token Pruning and Geometric Priors

Many recent Transformer-based 3D HPE(Human Pose Estimation) methods treat each frame of the video as an individual pose token and input it into the model for computation, which leads to significant computational overhead.Furthermore, these works lack a quantitative analysis of the depth ambiguity error in the process of dimensionality lifting from 2D to 3D, which resulting in suboptimal prediction accuracy.To address the aforementioned issues, this paper introduces a KNN-based adaptive token pruning method,which greatly reduces the number of tokens the model needs to process, thereby saving computational resources. Meanwhile, to tackle the issue of depth ambiguity, we leverages the pin-hole imaging principle and employs geometric prior knowledge to decouple the 2D-to-3D lifting mapping, improving the model's accuracy. Our model TPGP(token pruning and geometric priors) achieves faster inference speeds and superior performance on the mainstream 3D human pose datasets.