Model's non-deterministic behaviour: Model produces different bounding box coordinates for the same image

vision
1

Hi all,

I have a ResNet-50 model trained (using detectron2 framework) to detect just 1 object.
The problem is for the same image the model will produce slightly different bounding box values:
#1

({'pred_boxes': Boxes(tensor([[832.2043, 473.2896, 986.7280, 521.5142]])),
  'scores': tensor([1.0000]),
  'pred_classes': tensor([0])},
 array([[0.99999404]], dtype=float32))

#2

({'pred_boxes': Boxes(tensor([[832.2321, 473.2645, 986.7278, 521.4969]])),
  'scores': tensor([1.0000]),
  'pred_classes': tensor([0])},
 array([[0.99999404]], dtype=float32))

I’d appreciate your guidance, as I’m very new to this.

2

Can you describe what kind of problems this causes for you? This is essentially a relative error in the 1e-4/5 range, which I would not worry about unless I absolutely have to.

Best regards

Thomas