Spatial Grounding
A task requiring a model to give the location of a target within a frame, typically expressed as a bounding box. Evaluation commonly uses IoU or accuracy. It complements temporal grounding by answering where the target is located in the image.
Timeline (last 90 days)
Each track of OmniCompass contains three types of complementary tasks: temporal grounding (TG), spatial grounding (SG), and question answering (QA).
Even on leading models, precise temporal and spatial localization remains a core unsolved challenge; models perform acceptably on coarse-grained semantic understanding, but have clear gaps on precise localization tasks.
All Facts (2)
Each track of OmniCompass contains three types of complementary tasks: temporal grounding (TG), spatial grounding (SG), and question answering (QA).
50%UnverifiedEven on leading models, precise temporal and spatial localization remains a core unsolved challenge; models perform acceptably on coarse-grained semantic understanding, but have clear gaps on precise localization tasks.
50%