Pose detection: Integrating machine learning with large vision models

Abstract

This paper presents a novel framework for automated yoga pose analysis that integrates computer vision with large Vision Models (LVMs) to provide detailed assessment and personalized feedback. Our system leverages Detectron 2 for pose detection and Qwen VL 2 for comprehensive pose evaluation, creating a pipeline that can identify misalignments and generate actionable guidance comparable to human instructors. In controlled evaluations across multiple yoga pose categories, our approach demonstrated superior joint accuracy (0.912) compared to established frameworks like MediaPipe (0.891) and AlphaPose (0.875). Most notably, complex poses such as inversions and deep twists showed the greatest differential benefit (53.2% improvement). Our findings demonstrate that the integration of advanced pose detection with vision-language models creates a synergistic effect that significantly enhances yoga learning outcomes. This work establishes a foundation for intelligent assistive systems in physical practice domains where precise form and alignment are critical for both effectiveness and safety.