LoadAnnotations and Metrics Summary issue

[Ihssene-Brahimi]

Hello,

1. When running UltraSAM in inference mode, I noticed that the LoadAnnotations transform failed to load the box annotations from the coco json file. Could you please clarify that ? The issue persisted until i implemented a separate AnnotationLoader Class which eventually worked.
2. I tested the model on the BUSBRA dataset and I got 0.0 everywhere (see image below) for the benign class and very poor metrics for the malignant class as well. The predictions were good enough when i plotted them, but the metrics did not reflect them, why?

Thanks !

[Chop1]

Hi,

I would advise checking if you have specified the metadata in the config (classes) and that your coco.json has the category ID starting at 1 (and not 0).

metainfo = {'classes': ('breast_nodule_benign', 'breast_nodule_malignant')}

test_dataloader = dict(
    dataset=dict(
        data_root=data_root,
        metainfo=metainfo,
        data_prefix=dict(img='BUSBRA/images'),
        ann_file='BUSBRA/annotations/test.BUSBRA__coco.json',
    ),
)

Hope this helps!

[hmrasa]

Hi, great job!
Is there any sample script that takes an image as input and generates the segmentation mask?