Object Detection : Training with images that do not contain objects

Hello,

I am training a Mask RCNN model. Some of my images in the training data do not have any objects I want to detect.

According to:

https://de.mathworks.com/help/vision/ref/trainmaskrcnn.html

https://de.mathworks.com/help/vision/ug/getting-started-with-mask-r-cnn-for-instance-segmentation.html

I need a Mx4 matrix for the bounding boxes for each object.

What value should I past as matrix and label if the training image does not contain any object?

Should I leave these images out?

Best Regards

[Wai Hin Cheung]

Have you found a solution to this?

I am also facing the same problem.

What happens if I have an image that does not contain any object? When I feed it into the trained model, will the detector be able to output something like 'no object detected in this image'?

[Odo Luo]

Almost forgot I posted this question - but yes, I think so.

Mask RCNN work with image proposals and has an additional category 'background' build into the model. I just left the images without object out. The images with objed has enough "background" to be trained on.

Answer 1

[Odo Luo]

Mask RCNN takes random proposals during training. So its fine