Yu Siyue ; Xiao Jimin ; Zhang Bingfeng ; Lim Eng Gee

Abstract

Co-salient object detection, with the target of detecting co-existed salientobjects among a group of images, is gaining popularity. Recent works use theattention mechanism or extra information to aggregate common co-salientfeatures, leading to incomplete even incorrect responses for target objects. Inthis paper, we aim to mine comprehensive co-salient features with democracy andreduce background interference without introducing any extra information. Toachieve this, we design a democratic prototype generation module to generatedemocratic response maps, covering sufficient co-salient regions and therebyinvolving more shared attributes of co-salient objects. Then a comprehensiveprototype based on the response maps can be generated as a guide for finalprediction. To suppress the noisy background information in the prototype, wepropose a self-contrastive learning module, where both positive and negativepairs are formed without relying on additional classification information.Besides, we also design a democratic feature enhancement module to furtherstrengthen the co-salient features by readjusting attention values. Extensiveexperiments show that our model obtains better performance than previousstate-of-the-art methods, especially on challenging real-world cases (e.g., forCoCA, we obtain a gain of 2.0% for MAE, 5.4% for maximum F-measure, 2.3% formaximum E-measure, and 3.7% for S-measure) under the same settings. Code willbe released soon.