Enhanced human pose estimation via self-distilled and token-pruned transformer
Abstract
Human pose estimation (HPE) is a fundamental challenge in computer vision, aiming to detect anatomical keypoints in images. Traditional methods rely on CNN models, but recent advancements in Vision Transformer (ViT) models have shown superior performance. However, ViTs often require substantial computational resources. This paper introduces SPTPose, a method that employs self-distillation and token pruning to reduce computational costs while maintaining high performance. Our SPTPose-B achieves a mAP of 74.8% on the MSCOCO validation set with only 13.2 million parameters and 4.7 GFLOPs. The source code is available at https://github.com/duduxx123/SPTPose .
Fields of Study & Subjects
Human Pose and Action Recognition
Gait Recognition and Analysis
Context-Aware Activity Recognition Systems
Publication details
- Journal / source
- Scientific Reports
- Publisher
- Springer Science and Business Media LLC
- Volume
- 16
- Issue
- 1
- Publication date
- 14 June 2026
- Language
- EN