AccLoc: Anchor-Free and two-stage detector for accurate object localization. (June 2022)

• Record Type: Journal Article
• Authors: Piao, Zhengquan; Wang, Junbo; Tang, Linbo; Zhao, Baojun; Wang, Wenzheng
• Is Part Of: Pattern recognition. Volume 126(2022)
• Publication Date: 2022-06
• Subjects: Object detection -- Accurate localization -- Anchor-free -- NMS-free -- Two-stage; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108523
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We proposed an anchor-free, NMS-free and two-stage object detection framework for accurate localization. Adaptive proposal jitter strategy contributes to learning the mappings between the input regional features and the desired output offsets. Proposal fine-tuning strategy encourages the networks to cover various and helpful features for the subsequent bounding box regression process. Extensive experiments on COCO, KITTI, and BDD100K were conducted to demonstrate the effectiveness of our proposed approach. Abstract: Current anchor-free object detectors have obtained detection performances comparable to those of anchor-based object detectors while avoiding the weaknesses of anchor designs. However, two challenges limit the localization performance. First, such anchor-free detectors have one stage that predicts the classification and localization results directly. A large regression space reduces the localization performance of such methods. Second, most of the existing detectors extract features which are ineffective for accurate localization. In this paper, for the first challenge, we propose two-stage networks to predict regression results stage by stage, thereby reducing the scope of the prediction space. For the second challenge, we design two novel modules with the aim of extracting effective features for accurate localization. Experimental results validate that each module in our approach is effective and validate that our approach has better object localization performance than previous related and advanced methods.