Co-salient Object Detection Guided by Category Labels

Acquiring pixel-level labels is laborious and time-consuming,whereas image-level labels can be obtained much more easily.However,the use of image-level labels for co-salient object detection(CoSOD) remains underexplored.This paper presents a two-stage approach for weakly supervised CoSOD,relying solely on image-level labels(class labels) for model training.By utilizing the semantic information of class labels,this approach enables the localization and segmentation of co-salient objects.In the first stage,a pseudo-label generation network is proposed to generate saliency maps for input images,supervised by class labels.In the second stage,a co-salient object segmentation network is trained using these saliency maps as pseudo-labels.A self-corrective learning strategy is also incorporated to enhance model performance.For the first time,this paper proposes using image-level labelsbased training approach for CoSOD.Experiments on three representative datasets demonstrate the effectiveness and feasibility of the proposed method.