Artefacts when segmenting an image using bbox xyxy

I've been stuck of figuring out the reason of why this is happening.

I'm segmenting by passing the SAM a xyxy format bbox.
bbox: [tensor(689.), tensor(29.), tensor(3728.), tensor(2915.)]

Can anyone confirm the artifacts by running the segmentation ?

What could have done wrong and is there any way to fix it ?

How I get the mask

```
mask_predictor = SAM2ImagePredictor(sam_model)
mask_predictor.set_image(im_rgb)
masks, scores, logits = mask_predictor.predict(
box=xyxy,
multimask_output=True,
)
```

Input image


Segmentation mask resulting from input image of dtype uint8


Segmentation mask resulting from input image of dtype float32

Even if it does not male much sense i tried passing the image with dtype=np.float32