C Concordresearch demos
CIDR 2027 · Interactive artifacts

Two video queries.
Two optimized rewrites.

Concord treats video as a relation, so a query optimizer can rewrite across modalities instead of feeding whole videos to a model. Each demo below exposes evaluated operator outputs step by step—no API key is required.

Browse the demos MMDS repository ↗
Evaluated case studies

Pick a query to explore

O1 + O2 · Modality substitution & temporal pushdown

Event localization

Explore lecture demonstrations and soccer goals through the same video-only baseline, transcript-only rewrite, and transcript-guided rewrite. Inspect how each operator transforms the source into localized intervals or timestamps.

Examples
Lecture + soccer
Alternatives
Baseline, O1, O2
Output
Intervals + points
Open event localization →
O3 · Detect → Track → Join

Cross-camera vehicle association

Associate vehicles across two synchronized highway cameras. Compare the semantic Reduce → Unnest query, which asks one model to discover and match vehicles directly, against the optimized rewrite whose detection, tracking, and join relations can be inspected one operator at a time.

Cameras
2 synchronized
Operators
Inspectable
Source
I-24 MOTION
Open cross-camera join →
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 ↗