Reproduce “Learning Non-target Knowledge for Few-shot Semantic 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
Yuanwei Liu, Nian Liu, Qinglong Cao, Xiwen Yao, Junwei Han, Ling Shao
NTRENet reduces false positives in few-shot semantic segmentation by modeling what should be excluded as well as what should match the target. It combines background removal, distracting-object suppression, and prototype-level contrastive learning, with reported gains over comparison methods in several one-shot benchmark settings.
Key insights
- Explicitly representing non-target regions can complement target prototypes and reduce false activations from both background and co-occurring objects.
- Distracting-object prototypes are useful hard negatives: the reported comparison found them more effective than negatives representing background alone or all non-target pixels.
- Each component contributes in the reported ablation, though gains in mIoU are less consistent in five-shot settings than in one-shot evaluation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.