Surface Normal

This is a toolbox used to compute surface normal from a depth map.

Method

Our surface normal method is mainly based on plane fitting. Therefore, not only dense but sparse are both covered in our case. As for proof of plane fitting, please refer to least square for details. Following parameters are required:

float focal length;
float cx, cy;
int WINDOWSIZE; // for nearby valid 3D points searching
float Threshold; // the max allowed depth difference between nearby 3D points

How to read the surface normal value from image

In our code, for better visualization, we convert the real surface normal (nx, ny, nz) to picture as follows:

// For 3 Channles in saved Image(imgx, imgy, imgz): 
1st Channel: imgx = (nx + 1) * 255.0 / 2.0
2nd Channel: imgy = (ny + 1) * 255.0 / 2.0
3rd Channel: imgz = (nz + 1) * 255.0 / 2.0

// To get the real surface normal from Image:
nx = imgx * 2.0 / 255.0 -1
ny = imgy * 2.0 / 255.0 -1
nz = imgz * 2.0 / 255.0 -1

Prerequisites

OpenCV

We use OpenCV to manipulate images and features. Dowload and install instructions can be found at: http://opencv.org. Required at leat 3.0. Tested with OpenCV 4.5.0.

Building surface normal toolbox

Clone the repository:

git clone https://github.com/youmi-zym/surface_normal

Build: we provide a script build.sh to build the toolbox.

cd surface_normal
chmod +x build.sh
./build.sh

NYUDV2 Example

./build/surface_normal nyudv2 ./asserts/depth.png ./asserts/sn_nyudv2.png 1000.0

More Example

for other datasets, such as KITTI2015, KITTI2012, ScanNet, we provide a script run.sh here. Refactor it to your style.

./run.sh

Acknowledgement

Code is mainly adopted from Cindy-xdZhang and JiaXiong Qiu: DeepLidar.

YifuTao / surface_normal