pytorch faster r-cnn, no bounding box image return value error(all bounding boxes should have positive height and width)

Asked Dec 14 '21 at 04:27

Active Dec 14 '21 at 06:08

Viewed 639 times

I'm trying to train Faster R-CNN on custom dataset. When I train with no object image(image without bounding boxes), it returns value error 'All bounding boxes should have positive height and width'.

My dataset contains image, target([xmin, ymin, xmax, ymax], label). For non-object image, target is ([0, 0, 0, 0], 0).

Is there any way to solve this problem?

I'm following this code. https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html

and which one is correct method for training non-object images?

use [0, 0, 0, 0, 0] for non-object image
use [0, 0, w, h, 0] for non-object image(w, h is size of original image)

edited Dec 14 '21 at 06:08

asked Dec 14 '21 at 04:27

Naeun Lee

Same problem on my side. Currently the only workaround I see is to give the bounding box a width and height of 1 pixel. However this would still influence training somehow. – Marcel_marcel1991 Dec 14 '21 at 10:22
I tried [0, 0, 0.5, 0.5] too. But, it returns 0 loss – Naeun Lee Dec 14 '21 at 10:32

pytorch faster r-cnn, no bounding box image return value error(all bounding boxes should have positive height and width)

0 Answers0