Error in random_transform of mask about keras-maskrcnn HOT 2 CLOSED

In addition to the error, I'm confused as to why the mask is set to be the the size of the image and why it already has an extra dimension.

The reason for the extra dimension is because some functionality of keras for transforming images requires 3 channel images. The reason for setting the mask size to be the shape of the image is because for the loss function we need it to be that shape. We don't know where in the image the network will predict a detection, providing the mask in the same frame as the image means we are prepared for any bbox. We could store the masks clipped, but that would mean we need to place the mask in an image frame somewhere in the generator. This is not a big deal of course, but if possible it's nice to prevent that. Additionally, storing it in the image frame won't give much data overhead, png for instance is perfectly capable of encoding big black spaces in a few bytes.

Regarding the error, let's continue the discussion on Slack so we can debug the problem. I'll report here when an issue is found so others can see it too if that is useful.

from keras-maskrcnn.

The issue was due to some of bounding boxes being out of my image. @hgaiser has since modified the code to account for this. Presently, it does so by ignoring those faulty annotations and not having the network train on them.

from keras-maskrcnn.