Receptive Field of View: The term is borrowed from biology where it describes the “portion of sensory space that can elicit neuronal responses when stimulated” (wikipedia). Each output pixel can look at/depends on an input patch with that diameter centered at its position. Based on this patch, the network has to be able to make a decision about the prediction for the respective pixel.
Early Stopping to avoid overfitting: define an EarlyStopping class
Three-class model (foreground, background, boundary),
Distance transform (label for each pixel is the distance to the closest boundary),
Edge affinity (consider not just the pixel but also its direct neighbors, predicts the probability that there is an edge, this is called affinity.) 听的时候懂了，回来看的时候没太看懂
Metric learning (learns to predict an embedding vector for each pixel.)