Data augmentation resizing bounding boxes

[ljtruong]

Has anyone tried implementing a custom augmentation tool such as albumentations?

When I transform the image with resizeTransform it is the same as albumentations resize.SmallestMaxSize function. However the bounding boxes returned from detectron2 are different.

original input

bboxes = [[2039, 1058, 2887, 1929], [436, 38, 1463, 1112]]
original image 'height': 4240, 'width': 3480

detectron

TransformList[ResizeTransform(h=4240, w=3480, new_h=975, new_w=800, interp=2)]
returns
'bbox': array([468, 243, 800, 686])
'bbox': array([100, 8, 436, 264])

albumentations

[(468, 243, 663, 443), (100, 8, 336, 255)]

Investigating the calculation here indicate the calculation for albumentations is number is correct. I may be missing something here.

calcualtions for the first bounding box
2039 * (800 * 1/3480) = 468
1058 * (975 * 1/4240) = 243
2887 * (800 * 1/3480) = 663
1929 * (975 * 1/4240) = 443

[ljtruong]

It seems I accidentially set my BoxMode incorrectly.