test

[lmr2706]

Hello, I also want to ask you a question. In general image segmentation tasks, ground truth is not supposed to be provided during the testing phase, right? However, in interactive segmentation, to simulate user interaction scenarios, we have to use the ground truth to generate initial interaction prompts during testing. Is this approach widely accepted? How should we handle data without ground truth in the real world?

[qinliuliuqin]

@lmr2706 Hi, ground truth is required for general image segmentation tasks during testing if you want to evaluate the segmentation performance (e.g., IoU). For interactive segmentation, we also use the ground truth to simulate the interactions that otherwise should be provided by users if the ground truth does not exist. This is the widely adopted evaluation protocol for interactive segmentation.

[lmr2706]

Hello, I have one more question. Do you have specific code for your text prompts? How are the text attributes obtained? For medical datasets where lesions are difficult to describe, do you have any good methods to obtain text attributes? Thank you for your explanation.