D-DL4CV-Lec17b-Surface_Normals

What is Surface Normals?

For each pixel in the image, surface normals give a vector giving the normal vector to the object

---

Predicting Surface Normals

Method

We use an RGB image as input, then predict the surface normal using fully convolutional network

Loss

Let $x$ be the predicted vector for a pixel, and $y$ be that of the ground truth, then the loss for this pixel is

$$L=\frac{x\cdot y}{|x||y|}$$