Reproduce “Glocal Energy-based Learning for Few-Shot Open-Set Recognition” 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
Haoyu Wang, Guansong Pang, Peng Wang, Lei Zhang, Wei Wei, Yanning Zhang
The Glocal Energy-based Learning (GEL) framework addresses Few-Shot Open-Set Recognition by integrating both global (class-wise) and local (pixel-wise) information. This dual approach enables the model to distinguish between known and unknown classes more effectively, achieving improved AUROC and F1 scores while maintaining competitive classification accuracy on datasets like miniImageNet, tieredImageNet, and CIFAR-FS.
Key insights
- Combining both global (class-wise) and local (pixel-wise) information is crucial for robust Few-Shot Open-Set Recognition (FSOR), especially when unknown classes share global semantic similarities with known classes.
- An energy-based module can effectively fuse global and local similarity scores to create a unified confidence measure, leading to better separation between known and unknown classes.
- The pixel-wise branch, using a k-nearest neighbor approach on feature maps, captures fine-grained local patterns that are essential for distinguishing subtle differences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.