Tag
SOCO benchmark evaluates structured object understanding in vision models through consistent part-level annotations and keypoint descriptions, revealing gaps between language-grounded localization and visual correspondence while demonstrating strong prediction of downstream task performance.