Reproduce “ANTS: Adaptive Negative Textual Space Shaping for OOD Detection via Test-Time MLLM Understanding and Reasoning” 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
Wenjie Zhu, Yabin Zhang, Xin Jin, Wenjun Zeng, Lei Zhang
ANTS improves zero-shot out-of-distribution image detection by adapting negative text labels to images observed in the test stream. It mines likely negative images, uses a multimodal language model to generate descriptions and labels resembling relevant known classes, then adaptively combines their scores. On OpenOOD with ImageNet-1K, it reports FPR95 of 60.98 near and 15.38 far, lower than listed competitors in this setup.
Key insights
- The test-time understanding and reasoning capabilities of MLLMs can be effectively harnessed to create highly descriptive and targeted negative textual spaces for robust OOD detection.
- Dynamically identifying and targeting relevant in-distribution (ID) classes for generating 'visually similar' negative labels is crucial for improving near-OOD detection and reducing false negatives.
- An adaptive weighting mechanism, informed by the observed performance of different negative textual spaces on mined negative samples, allows for seamless OOD detection across diverse scenarios (near and far-OOD) without explicit prior knowledge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.