Bad case
Cjl-MedSeg opened this issue · 8 comments
This Issue is used to collect some bad cases, which will help us iterate the model.
When i tried to crop out implants from XRay images , SAM Med2D fails miserably. SAM gives better results. ( SAM also having issues with many images. but it is far better ).
I am using bounding box and do segmentation.
SAM Output
SAM MED 2D prediction
This is one of the SAM ouput. I suppose this is the problem you tries to solve. I am much curious about MED2D in this context. but the output of SAM-MED2D is not convincing.
SAM MED2D prediction
Thank you for sharing the bad case. You can try predicting the results without using the adapter layer, which may slightly improve the performance. If the dataset is open source, would it be possible for you to provide us with information about the dataset? We would like to optimize our model for such data in the future.
The segmentations are generated with the bounding boxes provided by the authors and checkpoint "medsam_vit_b.pth", showing accurate results and implying that our local implementation is correct.
The segmentations are generated with the bounding boxes customized by the users and checkpoint "medsam_vit_b.pth", showing degraded results.
Those observations might demonstrate the unstable performance of the pre-trained model.
2 pts test
SAM-Med2D 1 pts
SAM-Med2D 2 pts
SAM web demo 2pts
in this case SAM's 2pts is better than SAM-Med2D
5 pts test
there are the five points
SAM-Med2D web demo output
SAM-B in SAM-Med2D's web demo
SAM web demo
the SAM web demo doesn't seem so bad
5 pts test there are the five points
SAM-Med2D web demo output
SAM-B in SAM-Med2D's web demo
SAM web demo
the SAM web demo doesn't seem so bad
An interesting comparison between SAM and its fine-tuned medical variant.
It seems that the pretrained model sam-med2d_b.pth works bad on Spine X-ray to segment vertebral body.