Train faster rcnn with negative images

Hi everyone,

I’m training a faster RCNN to detected object wich appear sometime in my scene, the detection works well but when the object doesn’t appear, I have a lot of false detection with high accuracy.

So my idea was to add images without object to detect (empty label file), so the faster RCNN learn the empty scene.

But when i’m training the following message is displayed:

1218/1334 [==========================>...] - ETA: 1:36 - rpn_cls: 0.0456 - rpn_regr: 0.0236 - detector_cls: 0.1179 - detector_regr: 0.1013
No GT bboxes found in my_image.png

Is it only a warning or is the image without ground true boxes is ignored so it’s useless to add images like this one?

Thank you

Hi steventel,
If your label file has no bboxes info, log will show “No GT bboxes found in …” .
Then skip this image.

Hello everyone,

I would like to reword the original question in order to be sure that I understand the behaviour of the library.

If my dataset contains 1000 images with some boxes in each image (1000 non empty label files), and 500 images with no boxes (500 empty label files), I will get the warning message “No GT bboxes found in …” for the concerned 500 images.

But are the 500 images totally discarded from training, and the training with my 1500 images is in fact the same as a training with only the 1000 images ?

In other words, does the warning message mean that my 500 images will never be used in a batch in the training algorithm ?

Yes, your 500 images will not be used.