Question regarding MIL loss

[WeiChihChern]

In mask-auto-labeler/models/mal.py,

Line 420:
# Multiple instance learning Loss
loss_mil = self.mil_loss(student_seg_sigmoid, mask)

The mask in self.mil_loss(student_seg_sigmoid, mask) is from line 410, where mask = F.interpolate(x['mask'], size=(oh, ow), mode='bilinear', align_corners=False).reshape(-1, oh, ow)

I wonder does this mean we are not only using bbox but also ground truth masks during training? Thank you.

[WeiChihChern]

Is x['mask'] basically a tight bbox of 1s?

[hongyeob-kim]

@WeiChihChern

I'm not the author but If you check below lines in mask-auto-labeler/datasets/voc.py, x['mask'] is simply box shaped masking filled with 1.
Is this what you are looking for?

mask-auto-labeler/datasets/voc.py

Lines 95 to 104 in 8427b66

	# box mask
	mask = np.zeros((h, w))
	bbox = ann['bbox']
	x0, y0, x1, y1 = int(bbox[0]), int(bbox[1]), int(bbox[0] + bbox[2]), int(bbox[1] + bbox[3])
	mask[y0:y1+1, x0:x1+1] = 1
	data = {'image': img, 'mask': mask, 'height': h, 'width': w,
	'category_id': ann['category_id'], 'bbox': np.array([bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]], dtype=np.float32),
	'compact_category_id': self.cat_mapping[int(ann['category_id'])],
	'id': ann['id']}

[WeiChihChern]

@hongyeob-kim Good find, thanks.

[voidrank]

Hi @WeiChihChern @hongyeob-kim ,

Thanks for asking and answering this question. It’s not a good idea to use mask for box-shaped mask. I will put this in the readme and maybe change the name in the future.

Best