Learning Foreground Information Bottleneck for few-shot semantic segmentation

Abstract

Few-shot semantic segmentation aims to segment unseen classes with only a few annotated samples, which has great values for the real-world application in the wild. However, since the target class is treated as the background in the training, the network tends to extract much irrelevant nuisance factors, which results in the feature undermining problem for the target class. Consequently, it is difficult to produce an accurate segmentation map. To address this problem, in this paper, we apply the information bottleneck theory to few-shot semantic segmentation and propose the Foreground Information Bottleneck (FIB) module. Based on the support information, FIB module filters out the irrelevant information and promotes the foreground-related feature paradigms. Meanwhile, to solve the intractable mutual information and enable the end-to-end optimization of FIB module, we derive the Foreground Information Bottleneck Loss (FIBLoss) according to the inherent attribute of few-shot segmentation. Moreover, since there exists severe noise interference in the wild, we design a Target Information Refinement (TIR) block to further exploit discriminative cues of foreground. TIR block calculates the pairwise interaction and exploits the detailed information of the foreground object, which is beneficial to the feature refinement. Extensive experiments on two challenging datasets reflect the proposed FIB module significantly improves the performance of few-shot segmentation and delivers the state-of-the-art results.