Reproduce “Large-Scale Evaluation of Open-Set Image Classification Techniques” 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
Halil Bisgin, Andres Palechor, Mike Suter, Manuel Günther
Researchers at the University of Michigan-Flint and the University of Zurich conducted the first large-scale comparative evaluation of open-set image classification techniques on ImageNet, utilizing three distinct semantic similarity protocols. Their analysis demonstrates that combining training with negative samples and Entropic Open-Set (EOS) loss with post-processing methods like OpenMax or PROSER yields the most robust performance for recognizing unknown classes.
Key insights
- Incorporating negative samples during training substantially improves the performance of open-set classification algorithms.
- The Entropic Open-Set (EOS) loss consistently enhances the ability of post-processing algorithms to detect unknown samples.
- Hybrid approaches, specifically combining EOS training with post-processing methods like OpenMax or PROSER, achieve the best overall results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.