.png){kind=icon}
.png){kind=icon}
Return
ARPose: Anatomical relation-driven token pruning for human pose estimation
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-026-02310-0.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Token pruning is widely used to improve the efficiency of transformer-based human pose estimation. However, existing pruning methods often overlook key human anatomical priors, such as the varying scales of body parts and their structured spatial alignment, which are critical for precise token selection. To address this gap, we proposeAnatomical Relation-driven Pose (ARPose), a novel token pruning framework that integrates scale and structural alignment priors for accurate and efficient pose estimation. ARPose adopts a two-stage pruning strategy. In the first stage, we develop an Anatomical Scale-driven Module (ASM) to extract multi-scale features through scale-adaptive patching, where smaller patches capture fine details (e.g., hands, facial features) and larger patches model broader body structures (e.g., torso, limbs). In the second stage, we design an Anatomical Alignment-driven Group Attention (AAGA) module to group joints and tokens based on body anatomical relationships, dynamically modeling dependencies across these groups. Specifically, through a learnable relation matrix, a statistically optimal anatomical relation is derived as a structural prior, followed by input-dependent feature aggregation and modulation that adapts this fixed anatomical prior to the unique features of each input image. An integrated context gating mechanism further fuses semantic information. Together, ASM and AAGA enable progressive, anatomy-aware pruning that reduces computational overhead while preserving a rich representation. Extensive experimental results on benchmark datasets demonstrate that ARPose achieves superior performance over existing methods.
Keywords:
Deep learning
Transformer
Pose estimation
Token pruning
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K