Labeled data is often too costly to obtain in real-world settings, driving interest in adapting models without annotations. According to the abstract, vision-language models (VLMs) can provide pseudo-labels for unlabeled images, but this strategy may fail when the model's vocabulary does not include all classes present in the target domain.
The paper, 'Vocabulary-Incomplete Unsupervised Vision-Language Model Adaptation,' focuses on this exact problem. The abstract indicates that the proposed method adapts VLMs using pseudo-labels while handling the vocabulary mismatch. However, the provided abstract is truncated, so specific technical details are not available in this summary.
The work is significant because it addresses a practical limitation of VLM-based pseudo-labeling: incomplete label sets. By enabling adaptation under such conditions, it could extend the use of unsupervised methods to more diverse and niche domains.