Friday, 16 October 2026 · 6 min read
OmniCompass benchmark evaluates multimodal video grounding and understanding comprehensively
OmniCompass introduces paired text and text+image target specifications across vision-language and audio-visual tracks, assessing temporal grounding, spatial grounding, and question answering on 820 videos, revealing gaps in grounding, audio-visual integration, and cross-specification consistency.
Based on the paper OmniCompass: A Holistic Benchmark for Omni-Modal Video Grounding and Understanding by Xuyang Liu, Shihao Wang, Xinyu Wei, Minghan Li and 4 others
Challenges in evaluating comprehensive omni-modal video understanding
Multimodal large language models (MLLMs) are advancing rapidly in processing and reasoning over images, videos, audio, and text, enabling complex video understanding tasks that require identifying targets, grounding them in space and time, and integrating audio-visual evidence.
Existing video understanding benchmarks focus on limited aspects such as only visual or audio-visual evidence, use only text or text+image target specifications, or cover only some tasks like question answering or grounding, making it hard to systematically diagnose model capabilities across all relevant dimensions.
A holistic evaluation requires assessing models on when a target appears (temporal grounding), where it is (spatial grounding), and what happens (question answering), under both text-only and text+image target specifications, and across visual and audio-visual evidence, which existing benchmarks do not fully provide.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short