Generalized Few-shot Semantic Segmentation
This work addresses the challenge of adapting semantic segmentation models to novel classes with limited data, which is important for applications requiring quick adaptation, but it is incremental as it builds on existing few-shot segmentation methods.
The paper tackles the problem of few-shot semantic segmentation by introducing a new benchmark called Generalized Few-Shot Semantic Segmentation (GFS-Seg) to analyze model generalization, showing that previous methods fall short in this setting. It proposes Context-Aware Prototype Learning (CAPL), which improves performance by leveraging co-occurrence priors and dynamically enriching contextual information, achieving significant gains on Pascal-VOC and COCO datasets.
Training semantic segmentation models requires a large amount of finely annotated data, making it hard to quickly adapt to novel classes not satisfying this condition. Few-Shot Segmentation (FS-Seg) tackles this problem with many constraints. In this paper, we introduce a new benchmark, called Generalized Few-Shot Semantic Segmentation (GFS-Seg), to analyze the generalization ability of simultaneously segmenting the novel categories with very few examples and the base categories with sufficient examples. It is the first study showing that previous representative state-of-the-art FS-Seg methods fall short in GFS-Seg and the performance discrepancy mainly comes from the constrained setting of FS-Seg. To make GFS-Seg tractable, we set up a GFS-Seg baseline that achieves decent performance without structural change on the original model. Then, since context is essential for semantic segmentation, we propose the Context-Aware Prototype Learning (CAPL) that significantly improves performance by 1) leveraging the co-occurrence prior knowledge from support samples, and 2) dynamically enriching contextual information to the classifier, conditioned on the content of each query image. Both two contributions are experimentally shown to have substantial practical merit. Extensive experiments on Pascal-VOC and COCO manifest the effectiveness of CAPL, and CAPL generalizes well to FS-Seg by achieving competitive performance. Code is available at https://github.com/dvlab-research/GFS-Seg.