Open-Vocabulary Attribute Detection

Skibidi vibes in the lab as researchers dive into open-vocabulary attribute detection. The goal is to flexibly identify visual concepts using any text prompt, lowkey pushing the limits of classical object detection. With a dank benchmark covering 117 attributes across 80 object classes, they are ready to ship the future of vision-language models. A whopping 1.4 million annotations are in play, enabling a sigma-level evaluation of attribute detection. The paper introduces a baseline method, edging closer to understanding how models grasp attributes. Goon vibes all around as they explore this fresh territory.