Fig. 5.

An example of the structured surface prediction. The left image shows that a surface can be represented by the row index at each column and the surface is defined on the grid (generated by either graph methods or manual selection). The right image shows that the surface position can be a float value (thus the accuracy is not limited by the pixel resolution), and the exact position for each column can be inferred independently by the network output q_i’s. With q₁ and q₂ being directly computed in this case.