With augmentations, remove bbox if less than 25% of the box is visible ?

[ExtReMLapin]

Hello,

When training with augmentations, including the ones from Albumentation, do you know if there is a way to automatically remove the annotations if less than (for example) 25% of the original box is visible ?

For example, this kind of edge (literally) case :

[glenn-jocher]

@ExtReMLapin anything < 10% of original area should already be removed by the augmentation box candidates function area_thr=0.1 argument.

ultralytics/ultralytics/yolo/data/augment.py

Lines 384 to 390 in 8c8c439

	def box_candidates(self, box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16): # box1(4,n), box2(4,n)
	# Compute box candidates: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio
	w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
	w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
	ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps)) # aspect ratio
	return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr) # candidates

[ExtReMLapin]

Thank you for the quick answer Glenn.

Don't you think there is something not working then ? From the screenshot I posted, i'm pretty sure the left box (expected ratio = 1:1) -which looks more like a line than a box- is less than 10% of the original size.

Keep in mind I'm also using Albumentations augmentations.

[glenn-jocher]

@ExtReMLapin if the original object was larger then yes there is a problem, but I can't really assume that without seeing the original labels for this image.

If you find any bugs and can reproduce them in a common dataset like COCO128 please let us know, as this should be working exactly as indicated in the candidate function.

[ExtReMLapin]

Alright, will take a look

[glenn-jocher]

@ExtReMLapin BTW the other candidate criteria are width and height > 2 pixels and a max aspect ratio i.e. (h/w or w/h) of <100