Perhaps missing sigmoid during prediction

Hi, Thanks for providing the code. 

I noticed that during training, the output of segmentation model is normalized for thresholding (https://github.com/devalab/DeepPocket/blob/main/train_segmentation.py#L340). But this output in prediction pipeline is not normalized (https://github.com/devalab/DeepPocket/blob/main/segment_pockets.py#L139-L141), which makes 0.5 not a good threshold anymore. 