Reproduce “Multi-Level Correlation Network For Few-Shot Image Classification” 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
Yunkai Dang, Min Zhang, Zhengyu Chen, Xinliang Zhang, Zheng Wang, Meijun Sun, Donglin Wang
Researchers at Westlake University introduced a Multi-Level Correlation Network (MLCN) for few-shot image classification, integrating self-correlation, cross-correlation, and pattern-correlation modules to capture local and semantic image features. This method achieved improved accuracy, for example, reaching 65.54% (1-shot) and 81.63% (5-shot) on miniImageNet, and 77.96% (1-shot) and 91.20% (5-shot) on CUB-200-2011.
Key insights
- Background information can act as a detrimental 'shortcut' during training, leading to models that generalize poorly to unseen novel classes.
- Explicitly capturing multi-level feature correlations (intra-image locality, inter-image semantic alignment, and fine-grained structural patterns) leads to more discriminative and transferable image descriptors.
- Combining these three distinct correlation modules with an anchor-based classification loss effectively enhances the performance of metric-learning for few-shot image classification.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.