Evidential Robust Feature Learning for Generalized Few-Shot Segmentation

Generalized few-shot segmentation aims to recognize novel class objects with limited annotated samples while preserving the ability to segment previously learned base class objects. However, the inadequate training samples make it a challenging task to capture robust features which are both generally transferable for learning novel classes while retaining the capability to discriminate for representing base classes without incurring forgetfulness. To tackle this challenge, we introduce an evidential learning approach that simultaneously improves feature representation robustness and reduces model uncertainty. Specifically, we leverage both general-purpose features from segmentation-centric large pre-trained foundation models and low-level structural features with a physics-informed constraint to improve representations, especially for novel classes. Moreover, we impose our model with the capability of uncertainty estimation by evidential learning. The awareness of the model uncertainty greatly improves the performance in novel classes. The calibrated estimation of model uncertainty, in turn, helps to quantitatively assess the feature robustness for few-shot learning. We evaluate our methods across a diverse range of datasets encompassing natural image datasets, medical image datasets, and remote sensing datasets. The experimental results on the PASCAL VOC, MS COCO, ISIC, ISPRS, and Zurich Summer datasets consistently demonstrate the efficacy of the proposed method and achieve new state-of-the-art. The source code will be made available at https://github.com/liuweide01/ERFL.