BAC
Beyond Anonymous Captions: Grounding Character Identity in Video Captioning and Question Answering
A benchmark of 750 captioned clips and 3,000 person-centric questions from LSMDC v2 movies, measuring identity-aware video understanding via QA accuracy.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info bac
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
1 sourced entryEvery number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Constructed from 92 LSMDC v2 movies, this dataset features 750 manually verified clips and 3,000 person-centric questions with identity-linked bounding boxes. It evaluates how well models ground character identities in video narratives.
Scoring uses overall QA accuracy. The study finds that combining visual face boxes with textual coordinates yields the best performance, with the proposed BAC-8B model achieving 93.20% accuracy.