Debiased Negative Mining Improves Out-of-distribution Detection with Pre-trained Vision-Language Models

Researchers proposed a debiased negative-mining method for post-hoc OOD detection in pre-trained vision-language models.

The work corrects bias in heuristic negative-label selection with a theoretical approximation of the negative-label distribution.

It is then turned into a Monte Carlo sampling scheme using ID labels and unlabeled corpus data.

Experiments show stronger OOD detection performance and new state-of-the-art results across multiple benchmarks.