Vision-language models (VLMs), such as CLIP, have gained popularity for their strong open vocabulary classification performance, but they are prone to assigning high confidence scores to misclassifications, limiting their reliability in safety-critical applications. We introduce a training-free, post-hoc uncertainty estimation method for contrastive VLMs that can be used to detect erroneous predictions. The key to our approach is to measure visual feature consistency within a class, using feature projection combined with multivariate Gaussians to create class-specific probabilistic embeddings. Our method is VLM-agnostic, requires no fine-tuning, demonstrates robustness to distribution shift, and works effectively with as few as 10 training images per class. Extensive experiments on ImageNet, Flowers102, Food101, EuroSAT and DTD show state-of-the-art error detection performance, significantly outperforming both deterministic and probabilistic VLM baselines. Code is available atthis https URL.
View on arXiv