The combination of values across all channels determines a pixel's attributes. Each pixel is also related to the surrounding pixels: generally, if the Euclidean distance across all channels between two adjacent pixels is not smaller than a certain threshold, they can be regarded as belonging to the same class
Apply convolution over the channels,
Idea: two-stage inference. First build a building semantic segmentation algorithm to obtain buildings (0-1), then use a graph generation algorithm whose inputs are the points specified by the user and the segmented image