AgentPose: Progressive Distribution Alignment via Feature Agent for Human Pose Distillation

2025-01-14

AI Summary

Existing pose distillation methods overlook the capacity gap between teacher and student models, leading to inefficient knowledge transfer. To address this, we propose a progressive distribution alignment distillation framework based on feature agents: first, a learnable Feature Agent is constructed to model the teacher’s feature distribution; then, a multi-stage distribution matching strategy guides the student’s features to progressively align with the teacher’s distribution. This work is the first to integrate explicit feature distribution modeling with progressive alignment into pose distillation, effectively bridging the model capacity gap. Evaluated on COCO, our method significantly improves the AP of lightweight student models—especially under large teacher-student capacity disparities, achieving a +3.2 AP gain. These results validate both the effectiveness and generalizability of our distribution alignment strategy.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Distributed Machine Learning & Federated Learning
• Multiagent Systems: Distributed Problem Solving

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Web Mining and Content Analysis: Large pretrained models with web data
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Pose distillation is widely adopted to reduce model size in human pose estimation. However, existing methods primarily emphasize the transfer of teacher knowledge while often neglecting the performance degradation resulted from the curse of capacity gap between teacher and student. To address this issue, we propose AgentPose, a novel pose distillation method that integrates a feature agent to model the distribution of teacher features and progressively aligns the distribution of student features with that of the teacher feature, effectively overcoming the capacity gap and enhancing the ability of knowledge transfer. Our comprehensive experiments conducted on the COCO dataset substantiate the effectiveness of our method in knowledge transfer, particularly in scenarios with a high capacity gap.

Problem

• Knowledge Distillation
• Human Pose Estimation
• Model Capability Gap

Innovation

• Knowledge Distillation
• AgentPose
• Human Pose Estimation