VRA
Loading query…
CIDR 2027 · Interactive artifact
One event query.
Three ways to execute it.
See how Concord substitutes a source-aligned transcript for video—or uses the transcript to restrict expensive video inference to a small fraction of the original video duration.
Baseline
Video → Localize—O1
Transcript → Localize—O2
Transcript → View → Localize—O1 + O2 · Cross-modal rewrites
Example
Event localization
Loading example…
Published evaluation · Table 2
Less video, better localization
Values below are frozen aggregate results over the evaluated lecture workload, not measurements recomputed from the trace above.
| Query | Video | Precision | Recall | F1 | MLLM tokens | Cost | Time |
|---|
Read the research
MMDS repository ↗
Concord: A Video Relational Algebra for Cross-Modal Query Optimization
Sultan Muratbek, Charisse Ivana Yeung, Chanwut Kittivorawong, and Alvin Cheung.