Integrating NAS for Human Pose Estimation
V. Srilakshmi, Uday Kiran G, B Moulika, G S Mahitha, G Laukya, M Ruthick · Procedia Computer Science · 2025
Neural Architecture Search (NAS) technologies have become popular across various fields, allowing for the joint learning of neural network architectures and weights. However, most existing NAS methods are task-specific, focusing on optimizing a single architecture to replace human-designed networks while often neglecting domain knowledge. This paper introduces Pose Neural Fabrics Search (PoseNFS), a unique NAS framework that integrates domain knowledge via part-specific neural architecture search—a form of multi-task learning—for human posture estimation. PoseNFS utilizes a novel search space called Cell-based Neural Fabric (CNF), employing a differentiable search approach to facilitate learning at both micro and macro levels. By utilizing prior knowledge of human body structure, PoseNFS directs the search for part-specific architectures personalized to different body components, treating the localization of human key points as multiple disentangled sub-tasks. Experimental results on the MPII and MS-COCO datasets demonstrate that PoseNFS significantly outperforms a manually designed part-based baseline model and several state-of-the-art methods, validating the effectiveness of this knowledge-guided strategy.