DM-NAS: Dynamic-Fusion Multi-Branch Neural Architecture Search for Human Pose Estimation

High-resolution net-works, while adept at continuously integrating multi-resolution in-formation, tend to force the network to absorb redundant information, thereby increasing computational complexity. A single-path single neural architecture search approach is an effective solution to alleviate this problem. In this study, Researches introduce a novel NAS method based on a single-path one-step architecture, called DMNAS. DM-NAS adapts the network structure to determine the best deep fusion construction method. Extensive experiments demonstrate that DM-NAS performs well on the COCO human posture estimation dataset, improving ac-curacy while reducing the number of model parameters. DM-NAS achieves an accuracy of 73.7, showcasing a 1 % improvement compared to HRNet, while necessitating merely 13.1M Params equivalent to 35.4% of HRNet.