RGB-encoded Normals |
Needle Map |
Albedo Map |
|
View 1: No Albedo Mapping |
View 2: No Albedo Mapping |
View 1: Albedo Mapping |
View 2: Albedo Mapping |
RGB-encoded Normals |
Needle Map |
Albedo Map |
|
View 1: No Albedo Mapping |
View 2: No Albedo Mapping |
View 1: Albedo Mapping |
View 2: Albedo Mapping |
RGB-encoded Normals |
Needle Map |
Albedo Map |
|
View 1: No Albedo Mapping |
View 2: No Albedo Mapping |
View 1: Albedo Mapping |
View 2: Albedo Mapping |
The reconstructions worked more or less identically to the sample solution. The
lightning directions generated by my solution were a little bit off, but the
difference was less than .002. The difference was insignificant with respect
to the final result. I also had a bit of hard time chasing a bug when computing the
depth because of NaNs in the normals. Once I removed those, everything worked
perfectly.
I think the results were quite good even though this approach has some limitations.
The one dataset that was very poorly reconstructed was the gray sphere. The stand of
the sphere was not illuminated very well, so the only part that was reconstructed was
the sphere itself. The owl dataset had a slight problem with the eyes (notice the
side view of the owl). I guess this is due to the fact that normals estimate is not
good in the areas of dark pixels. One way to solve/improve this is to use a measure of
confidence by weighing the contsraints for dark pixels less havily than the constraints for
bright pixels as described in the Bells & Whistles section of the project. Unfortunately,
I didn't have a chance to implement any of the extra credit suggestions.
Other limitations of this method of reconstruction are related to the fact that
the method is based on the assumption that the object has a Lambertian surface and it also
requires lighting calibration using a chrome sphere in the scene which can be somewhat cumbersone.
One way to perhaps create the illusion of a surface being diffuse is using low-intensity lighting,
so that to reduce the chances of specular highlights. Of course, we have to be careful that
the object is not too dark, otherwise the normal estimates will not be good.