OmniCompass-VL
An evaluation track within OmniCompass aimed at vision-language models, in which all task evidence is drawn from visual information. It comprises three task categories: temporal grounding, spatial grounding, and question answering.
Timeline (last 90 days)
OmniCompass is a holistic benchmark for omni-modal video localization and understanding, comprising two parallel evaluation tracks, OmniCompass-VL and OmniCompass-AV.
In OmniCompass-VL, question answering covers 5 perception categories and 3 reasoning categories; in OmniCompass-AV, question answering covers 3 perception categories and 2 reasoning categories.
OmniCompass-VL contains 820 videos and 2460 instances, covering multiple domains and three duration types: short, medium, and long.
All Facts (3)
OmniCompass is a holistic benchmark for omni-modal video localization and understanding, comprising two parallel evaluation tracks, OmniCompass-VL and OmniCompass-AV.
50%UnverifiedIn OmniCompass-VL, question answering covers 5 perception categories and 3 reasoning categories; in OmniCompass-AV, question answering covers 3 perception categories and 2 reasoning categories.
50%UnverifiedOmniCompass-VL contains 820 videos and 2460 instances, covering multiple domains and three duration types: short, medium, and long.
50%