Reproduce “Object-level Correlation for Few-Shot Segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chunlin Wen, Yu Zhang, Jie Fan, Hongyuan Zhu, Xiu-Shen Wei, Yijun Wang, Zhiqiang Kou, Shuzhou Sun
A new method for Few-Shot Semantic Segmentation (FSS) introduces an object-level correlation mechanism, addressing "hard pixel noise" by first mining general objects in query images and then correlating them with support targets. This approach achieves state-of-the-art performance on PASCAL-5i and COCO-20i datasets, improving mIoU by approximately 1% on PASCAL-5i with a ResNet-50 backbone.
Key insights
- Inspired by biological vision, segmenting a target object is more effective when done within a set of pre-selected salient "general objects" rather than from an entire image.
- Explicitly mining query "general object features" before constructing a target correlation effectively suppresses hard pixel noise, including irrelevant novel objects.
- Incorporating both foreground and background prototypes in the correlation construction module significantly enhances target discrimination and noise suppression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.