FCN Prediction Pixels don't make sense

[eyildiz-ugoe]

After getting no help on the Mask-RCNN, I wanted to give a try to FCN semantic segmentation again with my coco-style dataset. Trained the network for 20 epochs, 64 image size and with 3 classes (background + 2 classes), however, received the following result, which makes absolutely no sense whatsoever.

I am not sure what the problem is in both implementations in the repository, I guess they were tested only with shapes dataset, and they don't work with others. Or perhaps the loading function does not generate the masks properly.

Any idea?

[waspinator]

hmm, I'm having some trouble with FCN as well. I'm working on it.

Mask-RCNN should work though. I tested that on other datasets and it worked well. If your dataset is open and available, I can take a quick look.

[eyildiz-ugoe]

Sure, I already shared my dataset in the other post. See here.