Here we present better segmentation result using a modification of the UNET network published in version 1.5 of SONY Neural Network Console. Training images are prepared using the algorithm described here. Images were visually inspected and corrected, where necessary, using image editor.
The network was taken from NNC project “unetlike_125px_person.sdcproj”, included in the semantic segmentation example folder. Modification of the network were simple: changing the input parameters to accept 320×320 pixel images, and including some additional padding, where it was necessary.
Network architecture and some results are presented below: