Regulating confidence by corner discrepancy and center score in corner-based object detection methods

Article type: Research Article

Authors: He, Zihang | Zhao, Kaiyan | Li, Bohan | Li, Yong^{; *}

Affiliations: School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing, China

Correspondence: [*] Corresponding author. Yong Li. E-mail: [email protected].

Abstract: This paper proposes an approach that regulates the confidence of predicted boxes for corner-based detection methods. Corner-based methods have achieved state-of-the-art performance on MS-COCO by predicting corners and grouping them to generate boxes. However, the box confidence is simply defined to be the average score of grouped corners, ignoring the score and tag discrepancy between them. The discrepancy may lead to the generation of more false positives (FPs) since a larger discrepancy often means that the grouped corners less likely belong to the same object. Observing this, this paper proposes introducing the discrepancy of corners (DoC) to decrease the box confidence. Also, the score and location of center (SLoC) of a detection box is utilized to further finely regulate the confidence. DoC and SLoC can effectively reduce FPs and missings and hence improve the detection performance without changing any model parameter. Experimental results on MS-COCO also show improvements.

Keywords: Object detection, anchor-free, corner-based

DOI: 10.3233/JIFS-212804

Journal: Journal of Intelligent & Fuzzy Systems, vol. 44, no. 6, pp. 10709-10720, 2023