GRAC
The official government-established game content rating body in South Korea, responsible for legally mandated content rating reviews of game products distributed in the country. Its ratings carry legal authority.
Timeline (last 90 days)
GRAC's effectiveness depends heavily on the quality and coverage of the retrieval library, and there may be a lack of examples to draw on for rare surgical scenes; the graph construction process requires accurate instrument and tissue recognition, and end-to-end practicality still needs further validation.
On the EndoVis-18 benchmark, GRAC achieves a CIDEr-D score of 5.88, SGT achieves 4.23, and the strongest traditional Transformer baseline achieves 3.36.
On the factual consistency metric of the EndoVis-18 benchmark, GRAC's instrument-action-target triplet set accuracy reaches 52.6%, SGT reaches 32.7%, and the strongest traditional baseline reaches 26.2%.
On the TORS dataset, GRAC achieves a CIDEr-D score of 5.06, and the strongest baseline achieves 3.80.
GRAC (Graph-Based Retrieval-Augmented Captioning) is a framework that represents surgical scenes with graph structures and combines retrieval-augmented generation to improve caption quality.
GRAC models the surgical scene as a graph, where nodes represent instruments, actions and anatomical structures, and edges represent the interaction relationships among them.
The GRAC workflow has two steps: given an input scene graph, it first retrieves relevant "graph-description" paired samples from the training data; then it compares the retrieved samples with the instrument-tissue interactions in the input scene, uses contrastive features to adjust the retrieved description representations, and finally generates a textual description for the current scene.
In GRAC, the retrieval library consists of "graph-description" pairs from the training data, and the retrieval results are corrected by contrastive features before being used to generate descriptions of the current surgical scene.
When submitting the next version of their app, developers can provide their GRAC Rating Classification Number (RCN) to override the app's rating to any one of the four region-specific levels.
GRAC is South Korea's official game content rating body, and its ratings carry legal force.
1 more timeline events
All Facts (11)
GRAC's effectiveness depends heavily on the quality and coverage of the retrieval library, and there may be a lack of examples to draw on for rare surgical scenes; the graph construction process requires accurate instrument and tissue recognition, and end-to-end practicality still needs further validation.
50%UnverifiedOn the EndoVis-18 benchmark, GRAC achieves a CIDEr-D score of 5.88, SGT achieves 4.23, and the strongest traditional Transformer baseline achieves 3.36.
50%UnverifiedOn the factual consistency metric of the EndoVis-18 benchmark, GRAC's instrument-action-target triplet set accuracy reaches 52.6%, SGT reaches 32.7%, and the strongest traditional baseline reaches 26.2%.
50%UnverifiedOn the TORS dataset, GRAC achieves a CIDEr-D score of 5.06, and the strongest baseline achieves 3.80.
50%UnverifiedGRAC (Graph-Based Retrieval-Augmented Captioning) is a framework that represents surgical scenes with graph structures and combines retrieval-augmented generation to improve caption quality.
50%UnverifiedGRAC models the surgical scene as a graph, where nodes represent instruments, actions and anatomical structures, and edges represent the interaction relationships among them.
50%UnverifiedThe GRAC workflow has two steps: given an input scene graph, it first retrieves relevant "graph-description" paired samples from the training data; then it compares the retrieved samples with the instrument-tissue interactions in the input scene, uses contrastive features to adjust the retrieved description representations, and finally generates a textual description for the current scene.
50%UnverifiedIn GRAC, the retrieval library consists of "graph-description" pairs from the training data, and the retrieval results are corrected by contrastive features before being used to generate descriptions of the current surgical scene.
50%UnverifiedWhen submitting the next version of their app, developers can provide their GRAC Rating Classification Number (RCN) to override the app's rating to any one of the four region-specific levels.
50%UnverifiedGRAC is South Korea's official game content rating body, and its ratings carry legal force.
50%UnverifiedThe first change takes effect immediately: if a developer's app has received an official rating from the Game Rating and Administration Committee (GRAC) of Korea, that official rating can be used to override the app's existing automatic age rating on the South Korean App Store.
50%