RA-YOLOX: Re-parameterization align decoupled head and novel label assignment scheme based on YOLOX

Abstract

YOLOX is a state-of-the-art one-stage object detection model for real-time applications that employs a decoupled head and advanced label assignment. Despite its impressive performance, YOLOX has limitations that prevent it from achieving optimal accuracy in real-time settings. To improve these limitations, we propose a new approach called re-parameterization align YOLOX (RA-YOLOX). Our approach employs a novel re-parameterization align decoupled head to align the classification and regression tasks, enhancing the learning of connection information between classification and regression. In addition, we propose a novel label assignment(LA) scheme that effectively defines positive and negative samples and precisely designs loss weight function. Our LA scheme enables the detector to focus on high-quality positive samples and filter out low-quality positive samples during training. We provide three sizes of lite models, namely RA-YOLOX-s, RA-YOLOX-tiny, and RA-YOLOX-nano, all of which outperform YOLOX models of similar size by an average precision of 2.3%, 1.5%, and 1.7%, respectively, on the MS COCO-2017 validation set, demonstrating the efficacy of our approach. Our code is available at github.com/hcmyhc/RA-YOLOX.