ComCa has been accepted to CVPR

This research leverages pre-trained vision-language models (VLMs) for the task of attribute detection, which involves identifying all attributes of an object, such as color, material and texture.

ComCa, our training-free method, surpasses all other training-free models and is even competitive with training-based approaches.

By introducing a caching strategy that stores examples of objects with specific attributes, ComCa can refine the VLM's predictions at test time requiring no specific training nor manually annotated data. Indeed, the cache is constructed automatically by leveraging web-scale databases, such as CC12M, and large language models (LLMs) such as GPT by OpenAI.

Code will be released shortly, and the camera-ready version of the paper will be uploaded on arXiv as soon as it is ready.