GRAC: Graph-Based Retrieval-Augmented Captioning for Surgical Scene Description
arXiv cs.AI · 7 Nov 2026, 04:47 AEST
GRAC is a new framework designed to automate surgical scene descriptions by using graph-based representations. It shows significant improvements in language quality and factual agreement compared to existing methods.
Key facts
- GRAC (Graph-Based Retrieval-Augmented Captioning) is a framework that represents surgical scenes as graphs capturing relationships among instruments, actions, and anatomical structures.
- On the EndoVis-18 benchmark, GRAC achieved a CIDEr-D of 5.88, compared with 4.23 for SGT and 3.36 for the strongest conventional Transformer baseline.
- Exact instrument-action-target triplet-set accuracy reached 52.6%, compared with 32.7% for SGT and 26.2% for the strongest conventional baseline.
- On TORS, GRAC achieved a CIDEr-D of 5.06, compared with 3.80 for the strongest baseline.
Read aloud
GRAC is a new framework designed to automate surgical scene descriptions by using graph-based representations. It shows significant improvements in language quality and factual agreement compared to existing methods. GRAC Graph-Based Retrieval-Augmented Captioning is a framework that represents surgical scenes as graphs capturing relationships among instruments, actions, and anatomical structures. On the EndoVis-eighteen benchmark, GRAC achieved a CIDEr-D of five point eight eight, compared with four point two three for SGT and three point three six for the strongest conventional Transformer baseline. Exact instrument-action-target triplet-set accuracy reached fifty two point six percent, compared with thirty two point seven percent for SGT and twenty six point two percent for the strongest conventional baseline. On TORS, GRAC achieved a CIDEr-D of five point zero six, compared with three point eight zero for the strongest baseline.
Sources in this cluster
- GRAC: Graph-Based Retrieval-Augmented Captioning for Surgical Scene Description
arXiv cs.AI · 7 Nov 2026, 04:47 AEST