Abstract: Highlights•A new two-step normal estimation method.•Integrate geometric priors into deep learning framework.•Replace multi-scale architecture by multi-scale geometric input.•Achieve 10.79 angle error in comparison with the previous state of the art of 11.78.
Loading