Benchmarks

Six benchmarks. For each: what it contains, and one real record.

VALU

evaluated in this demo Benchmark evaluation tasks with their own annotations Paper · Annotations · Masked media

Temporal localization and understanding of video anomalies at five semantic levels.

AnnotationsADC multiple-choice records, TDG description + single time span, ALD description + single time span.

Videos fromUCF-Crime, MSAD, ECVA (ECVA extends CUVA).

not evaluated here Language annotation collection (question answering) Paper · Data

Large-scale question answering about surveillance footage.

AnnotationsQuestion–answer pairs.

Videos fromBuilt on existing surveillance video datasets; original videos are acquired separately.

not evaluated here Language annotation collection (structured) Project · Annotations

Fine-grained description of what happened, who was involved and where.

AnnotationsStructured events, entities and scene descriptions.

Videos fromSource-video lineage: not verified in this build.

not evaluated here Language annotation collection Paper

Surveillance video-and-language understanding: description and temporal grounding.

AnnotationsSentence descriptions with temporal annotations.

Videos fromUCF-Crime.

not evaluated here Benchmark with causal annotations Paper

Causation understanding of video anomalies: what happened, why, and how severe.

AnnotationsFree-text cause, effect and description annotations with time spans.

Videos fromVideos collected from online video platforms (VALU paper, related work).

not evaluated here Instruction-tuning data Paper · Code

Hierarchical instruction data used to train Holmes-VAU (clip, event and video level).

AnnotationsHierarchical instruction–response pairs.

Videos fromSurveillance anomaly datasets including UCF-Crime (Holmes-VAU paper).