matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.65k stars 11.7k forks source link

Negative training samples #1124

Open eyildiz-ugoe opened 5 years ago

eyildiz-ugoe commented 5 years ago

Hi,

I would like to implement a detector such as balloon detector which is given as an example here. I was wondering how could we even expect this to work without training on negative training images? Images that are not balloons (or whataver our target is) but nevertheless they should be trained on. I do not see this with mask_rcnn, and as a result, this balloon/whatever detector detects the objects that are not balloons/whatevers.

Is there a way around this? Am I missing something?

Thanks in advance.

keineahnung2345 commented 5 years ago

I guess it's because Mask R-CNN can learn what is background from the "zeros" in the masks. But if you want to include some images which don't contain any object, try this PR: https://github.com/matterport/Mask_RCNN/pull/1088.