一种基于点标注的弱监督目标检测方法

In recent years, weakly supervised object detection has attracted more and more researchers' attention in the field of computer vision and machine learning. Point annotation is one kind of weak annotation, which manually assigns an annotation on a point that belongs to an object. In the past few years, many weakly supervised object detection approaches based on deep learning have been proposed, but the current research on object detection based on point annotation is still blank. Considering that point annotation can provide abundant information about the location, category, and quantity of objects for weakly supervised object detection, this paper proposes a weakly supervised object detection algorithm based on point annotation. The algorithm makes up for the lack of supervision information in weakly supervised learning by exploring the dependency relationship between the annotation point and the object, between categories, and between instances. The proposed approach can enhance the performance of weakly supervised object detection. Three branches are introduced to enhance the learning process of weakly supervised object detection: spatial graph branch (SGB) uses the local correlation between the annotation point and the object to seek the relationship between point annotation and its spatial context; multi-semantics branch (MSB) uses the semantic co-occurrence probability between categories to construct the semantic typology and explore the global semantic relationship between annotations; count-guided instance branch (CIB) adopts the spatial local independence and feature differences between different instances to infer the pseudo supervision information of each object and achieve instance-level supervision. Training weakly supervised object detection approach with point annotation can save the cost of manual annotation and provide more abundant supervision information, which enhances the performance of weakly supervised object detection in essence. The experimental results on PASCAL VOC 2007 datasets and PASCAL VOC 2012 datasets display that compared with the baseline, the mAP of the proposed method is improved by 7.9% and 10.2% respectively, and the CorLoc is improved by 9.7% and 11.7% respectively.