Fig. 2.

Label vector encoding for a given image sample I with λ = 3 target surfaces. The indexes of vector positions are shown below each cell. The target label L(.) used in network training is the ordered concatenation of each surface encoded vector.