OmniCompass
A holistic benchmark for omni-modal video grounding and understanding, divided into two evaluation tracks for vision-language models and omni-modal models. It covers three task categories—temporal grounding, spatial grounding, and question answering—and is used to systematically diagnose the video understanding capabilities of multimodal large models.
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.
OmniCompass's core design philosophy is to require omni-modal video understanding to cover three dimensions simultaneously: when the target appears, where it is, and what happened, and to also examine the joint use of visual and auditory evidence.
Each track of OmniCompass contains three types of complementary tasks: temporal grounding (TG), spatial grounding (SG), and question answering (QA).
OmniCompass adopts a paired evaluation design: each instance is evaluated separately under two settings, text-only and text+image, to test whether the model truly understands the target reference rather than relying on the prompt form to take shortcuts.
All OmniCompass instances have undergone human annotation and multiple rounds of quality verification.
All Facts (5)
OmniCompass is a holistic benchmark for omni-modal video localization and understanding, comprising two parallel evaluation tracks, OmniCompass-VL and OmniCompass-AV.
50%UnverifiedOmniCompass's core design philosophy is to require omni-modal video understanding to cover three dimensions simultaneously: when the target appears, where it is, and what happened, and to also examine the joint use of visual and auditory evidence.
50%UnverifiedEach track of OmniCompass contains three types of complementary tasks: temporal grounding (TG), spatial grounding (SG), and question answering (QA).
50%UnverifiedOmniCompass adopts a paired evaluation design: each instance is evaluated separately under two settings, text-only and text+image, to test whether the model truly understands the target reference rather than relying on the prompt form to take shortcuts.
50%UnverifiedAll OmniCompass instances have undergone human annotation and multiple rounds of quality verification.
50%