Reducing Information Bottleneck for Weakly Supervised Semantic Segmentation

Weakly supervised semantic segmentation produces pixel-level localization\nfrom class labels; however, a classifier trained on such labels is likely to\nfocus on a small discriminative region of the target object. We interpret this\nphenomenon using the information bottleneck principle: the final layer of a\ndeep neural network, activated by the sigmoid or softmax activation functions,\ncauses an information bottleneck, and as a result, only a subset of the\ntask-relevant information is passed on to the output. We first support this\nargument through a simulated toy experiment and then propose a method to reduce\nthe information bottleneck by removing the last activation function. In\naddition, we introduce a new pooling method that further encourages the\ntransmission of information from non-discriminative regions to the\nclassification. Our experimental evaluations demonstrate that this simple\nmodification significantly improves the quality of localization maps on both\nthe PASCAL VOC 2012 and MS COCO 2014 datasets, exhibiting a new\nstate-of-the-art performance for weakly supervised semantic segmentation. The\ncode is available at: https://github.com/jbeomlee93/RIB.\n