Research

Abstract

Weakly supervised semantic segmentation aims to produce pixel-level localization fromclass labels. However, classifiers trained on such labels tend to focus only on small,discriminative regions of the target object. In this work, we interpret this limitationthrough the information bottleneck principle, where the final layer of a deep neuralnetwork, activated by sigmoid or softmax functions, restricts the flow of task-relevantinformation.We validate this phenomenon through a simulated toy experiment and propose a simple yeteffective solution by removing the final activation function to alleviate the informationbottleneck. Furthermore, we introduce a novel pooling method that encourages informationfrom non-discriminative regions to contribute to classification.Experimental results on PASCAL VOC 2012 and MS COCO 2014 demonstrate that our approachsignificantly improves localization quality and achieves state-of-the-art performancein weakly supervised semantic segmentation.The code is available at:https://github.com/jbeomlee93/RIB