Dear Theodore,
I am currently benchmarking language-guided medical image segmentation methods, particularly for referring expression segmentation.
A SAM-based pipeline using Grounding-DINO to generate bounding box proposals is an important and strong baseline. However, many papers do not clearly specify whether Grounding-DINO is fine-tuned on the target medical domain or evaluated directly in a zero-shot setting.
Fortunately, I found that Grounding-DINO was also compared in your paper, and BioMedParse demonstrated superior performance even without bounding box prompts.
I would like to ask for your advice: for a benchmark in medical image domains, should Grounding-DINO be fine-tuned on the target domain, or should it be evaluated in a zero-shot setting?
Thank you very much for any suggestions.
Dear Theodore,
I am currently benchmarking language-guided medical image segmentation methods, particularly for referring expression segmentation.
A SAM-based pipeline using Grounding-DINO to generate bounding box proposals is an important and strong baseline. However, many papers do not clearly specify whether Grounding-DINO is fine-tuned on the target medical domain or evaluated directly in a zero-shot setting.
Fortunately, I found that Grounding-DINO was also compared in your paper, and BioMedParse demonstrated superior performance even without bounding box prompts.
I would like to ask for your advice: for a benchmark in medical image domains, should Grounding-DINO be fine-tuned on the target domain, or should it be evaluated in a zero-shot setting?
Thank you very much for any suggestions.