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 lecture duration.

Explore the replay View published results

BaselineVideo → Localize—

O1Transcript → Localize—

O2Transcript → View → Localize—

O1 + O2 · Cross-modal rewrites

Lecture event localization

Find every interval where the lecturer creates a large, intact soap bubble that remains visible until it ruptures.

● recorded execution

VRA Loading recorded query…

Now inspecting

Loading…

Source videoMIT 8.03 · Lecture 20

129.8 s retained

0:0082:25

Processed video Reference events

Source-aligned representation

Transcript evidence

—

Materialized View

Candidate clip

—

Reference Prediction

Excerpt from MIT OpenCourseWare 8.03SC, Fall 2016. Candidate clip shown for research demonstration.

Operator output

Localized events

—

Published evaluation · Table 2

Less video, better localization

All three queries use the same workload and output contract. Values below are frozen publication results, not recomputed in the browser.

Video processed—O2 lecture query

Candidate recall—4 of 4 events covered

MLLM cost—relative to video-only

F1 score—at tIoU ≥ .3

Lecture event-localization results at the primary tIoU threshold.
Query	Video	Precision	Recall	F1	MLLM tokens	Cost	Time

Read the research

Concord: A Video Relational Algebra for Cross-Modal Query Optimization

Sultan Muratbek, Charisse Ivana Yeung, Chanwut Kittivorawong, and Alvin Cheung.

MMDS repository ↗

One event query.Three ways to execute it.