Abstract
Few-Shot Segmentation (FSS) is challenging for intra-class diversity and support sample scarcity. Many works focus on the class-wise or pixel-wise similarity between the support foreground and query sample while neglecting the support background, which is vital for FSS to suppress the related query background. In this paper, we propose a Multi-Similarity Enhancement Network (MSENet) to remedy this issue by extracting the pixel-wise support-query similarity of the foreground and background. To remedy the shift issue, caused by the huge difference between support and query target objects, we extract and fuse multiple support-query similarity, and keep enhancing them with convolutional operations. Experiments on three benchmarks demonstrate that our approach reaches SOTA performances and outperforms the best competitors with many desirable advantages.
Funding Information
  • National Key Research and Development Program of China (2020AAA0140004)