Open-vocabulary Attribute Detection

Abstract: Vision-language modeling has enabled open-vocabulary tasks where predictions can be queried using any text prompt in a zero-shot manner. Existing open-vocabulary tasks focus on object classes, whereas research on object attributes is limited due to the lack of a reliable attribute-focused evaluation benchmark. This paper introduces the Open-Vocabulary Attribute Detection (OVAD) task and the corresponding OVAD benchmark. The objective of the novel task and benchmark is to probe object-level attribute information learned by vision-language models. To this end, we created a clean and densely annotated test set cov- ering 117 attribute classes on the 80 object classes of MS COCO. It includes positive and negative annotations, which enables open-vocabulary evaluation. Overall, the bench- mark consists of 1.4 million annotations. For reference, we provide a first baseline method for open-vocabulary at- tribute detection. Moreover, we demonstrate the bench- mark’s value by studying the attribute detection perfor- mance of several foundation models.