Classification with an edge: Improving semantic image segmentation with boundary detection

Netwroks:

SEG-H encoder-decoder network

It’s a crossbreed of FCN and encoder-decoder architecture. Use pyramid-bottleneck architecture. Compared with SEG model, SEG-H combine the DEM and data as well in the database. For channels, except coulor channela(using pascal pre-trained model), it combines DSM and nDSM channel and initialized randomly using “Xavier” weight initialization which could make the gradient magnitude roughly the same across layers. These two streams are concatenated and fed through 1x1 convolution(linearly combines the vector of feature responses at each location into a score per class). Finally those scores are further converted to probabilities with a softmax layer.

HED-H multi-scale CNN

Add second branch for DSM. By using a regression los w.r.t. HED-H is mainly to detect the edge by height, and the color map HED-H is initialized by original HED model, and for height map it’s initialized by scratch.

FCN-N semantic segmentation network

It’s two FCN with initialzed by VGG and Pascal.

Conclude

This paper is mainly describe how to fuse several CNN together.