Get Started
Home
Topics
Search
Library
Research questionHow can zero-shot vision-language models adapt online to image corruption using only unlabeled test images?Zero-shot vision-language models can lose classification accuracy when weather, lighting, noise, or other corruption alters incoming images. Labeled data from each unexpected test distribution is costly to collect, so adaptation must occur from unlabeled test observations.
AI
Computer Vision
Image & Video Processing
Machine Learning
Multimodal Models
Research Paper
Latest papersRecent research connected to this question, newest first.Uniformity First: Uniformity-aware Test-time Adaptation of Vision-language Models against Image CorruptionThe source studies CLIP-style zero-shot classification under image-corruption shifts using unlabeled test data during inference, without collecting labeled data from the test distribution. Its evidence is experimental and reports accuracy results for the proposed UnInfo adaptation approach.research paper · Sep 2, 2026
Related questions
How can one vision model remain robust across changing and unseen adversarial perturbation budgets?How can image-poisoning attacks remain effective across supervised and self-supervised learning without obvious visual artifacts?How can zero-shot graph anomaly detectors remain reliable on unlabeled targets under domain shift?How can vision-language models maintain visual recognition when modalities are missing and source training data is unavailable?