文献类型：期刊文献

英文题名：Toward Minimal Misalignment at Minimal Cost in One-Stage and Anchor-Free Object Detection

作者：Hao, Shuaizheng[1]; Liu, Hongzhe[1]; Wang, Ningwei[1]; Xu, Cheng[1]

第一作者：Hao, Shuaizheng

机构：[1] Beijing Key Laboratory of Information Service Engineering, Beijing Union University, China

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2021

外文期刊名：arXiv

收录：EI(收录号：20210415510)

语种：英文

外文关键词：Alignment - Object recognition

摘要：Common one-stage object detectors consist of two sub-tasks: object classification and box localization, using two individual branches in head networks. The feature misalignment problem, caused by the different feature sensibilities between the two branches, hurts models' performance. In this letter, we rethink and investigate the problem in one-stage detectors and find the problem is further composed of scale-level misalignment and spatial-location misalignment specifically. Based on that observation, we propose Receptive Filed Adaptor (RFA), a simple and plug-in module for models' two branches, to augment each task's adaptability to different scale information. Further, we design a novel dynamic label assignment scheme, namely Aligned Points Sampler (APS), to dynamically mine the most spatially aligned feature points during the training procedure. The extensive experiments show that with a light cost, our proposal can consistently boost models' performance around 3 AP on MS COCO across different backbones. Our code is available at https://github.com/HaoGood/MOD. Copyright ? 2021, The Authors. All rights reserved.