Normal maps

[juanraul8]

Features:

• Clean tests (just change parameters in function call).
• Normal map computation.

Normal maps computed directly by depth maps:

Normals computed using point cloud normals:

Problem above was solved, see comments.

Future tasks:

• Experiment with bilateral filtering (checking against normal map)
• Pruning normal (exercise 3)
• Show normal map in unity.

[Zielon]

Why didnt you change computing normals in PointCloud? Can we fix it?

[juanraul8]

Why didnt you change computing normals in PointCloud? Can we fix it?

I do not think the problem is the normal computation. They are similar.

@BarisYazici normals visualization using PCL looks good.

First image is computing normals directly from depth maps (I follow a tutorial). Our normals where computed after backprojection, maybe that is the problem.

[juanraul8]

Ok, now I see what was the problem.

The problem was basically that m_normals does not have the grid structure. I also realize that normal computation does not seem to be same as paper formula.

Exercise 3 formula:

Kinect Fusion formula:

[juanraul8]

Another comment, in exercise 3, the following prune is applied:

const float du = 0.5f * (depthMap[idx + 1] - depthMap[idx - 1]);
                const float dv = 0.5f * (depthMap[idx + width] - depthMap[idx - width]);
                if (!std::isfinite(du) || !std::isfinite(dv) || abs(du) > maxDistanceHalved || abs(dv) > maxDistanceHalved) {
                    normalsTmp[idx] = Vector3f(MINF, MINF, MINF);
                    continue;
                }

Maybe we are missing something like that.

Paper mentioned a validity mask, which can be used to know if the vertex map or normal map pixel is valid.