MambaPose: A Human Pose Estimation Based on Gated Feedforward Network and Mamba
Jianqiang Zhang1, Jing Hou1, Qiusheng He1
1School of Electronic Information Engineering, Taiyuan University of Science and Technology, Taiyuan 030024, China.
Sensors (Basel, Switzerland)
|January 8, 2025
Summary
This study introduces a novel Mamba-based approach for human pose estimation, enhancing accuracy in crowded scenes and for small targets. The method improves keypoint detection robustness and anti-interference capabilities.
Area of Science:
- Computer Vision
- Machine Learning
- Deep Learning
Background:
- Human pose estimation is crucial in computer vision but struggles with dense crowds and small targets.
- Existing methods often exhibit false or missed detections in challenging scenarios.
Purpose of the Study:
- To develop an advanced human pose estimation technique using a Mamba-based architecture.
- To improve accuracy and robustness, particularly in complex environments with occlusions and dense crowds.
Main Methods:
- Designed a GMamba backbone network with a gating mechanism for precise keypoint localization.
- Implemented slice downsampling (SD) and multi-channel feature fusion for rich pose information.
- Introduced adaptive threshold focus loss (ATFL) to dynamically weight error-prone keypoints.
Main Results:
- Achieved an Average Precision (AP) of 72.2 and AP50 of 92.6 on the COCO 2017 validation set.
- Demonstrated a 1.1% improvement in AP50 compared to typical algorithms.
- Significantly enhanced accuracy and robustness in complex scenes, including cases of occlusion and intricate backgrounds.
Conclusions:
- The proposed Mamba-based human pose estimation method effectively detects human keypoints with improved accuracy and robustness.
- The GMamba backbone, SD, feature fusion, and ATFL contribute to superior performance in challenging scenarios.
- This approach offers a promising solution for reliable human pose estimation in real-world applications.
