In the same scenario, most of the reasoning can be correct, while a small portion yields no result.

[Shichang-Fu]

As shown in the image, the last two images in SAM's inference results correctly predicted the bounding box and mask. The first two images depict the same scenario, but predicted nothing.

I inferred from 500 images of similar scenarios, with a correct prediction rate of approximately 95% and a prediction rate of approximately 5% that predicted nothing.

I want to know why this happens and how to solve this problem.

[heyoeyo]

The first two examples both seem to work with either the default text prompt of visual or with pill. Here's the first one with prompt pill for example:

Maybe an error occurred when running those examples and that's why they didn't give a result?

[Shichang-Fu]

Yes, I was using "pill" as the prompt for inference, and about 5% of the images I got showed no target. I'm confused as to why, in the same scene, most images were detected correctly during inference, while a small portion lacked the target.

Additionally, I tried "visual" as the prompt. All images were segmented, but it wasn't as accurate as using "pill." "Visual" as the prompt also segmented the background.

[heyoeyo]

about 5% of the images I got showed no target

Maybe those images are blurry or otherwise not as clear as the other images that worked?