Ofir Lavi
CSEP576 – Computer Vision
|
|
|
|
Original |
RGB Normals |
Needle Map |
Albedo Map |
|
|
|
|
|
|
|
|
Reconstructed surface without Albedos |
Reconstructed surface with Albedos |
|
|
|
|
Original |
RGB Normals |
Needle Map |
Albedo Map |
|
|
|
|
|
|
|
|
Reconstructed surface without Albedos |
Reconstructed surface with Albedos |
|
|
|
|
Original |
RGB Normals |
Needle Map |
Albedo Map |
|
|
|
|
|
|
|
|
Reconstructed surface without Albedos |
Reconstructed surface with Albedos |
In general my solution gets the same exact results as the sample solution.
I tried two ways to calculate the lightning direction and both worked the same. Once with all 3 channels and once with only the Green channel (channel 1). I tested it with both techniques and the final results were the same. In order to reduce the shadows weight, I multiply both sides of the equation by (I) and the results were much better.
The solution will not work as well for shiny things, semi-translucent thing, inter-reflections and from looking at my Owl results, it will not work as well for non-reflective dark colors like the Owl eyes. An example for shiny thing can be seen in the owl belly, from some direction there appears to be a bump there while in realty there is not. A possible solution can be to paint the image with a non-shiny paint, but this can eliminate some of the features of the image.
We can see that the depth solution did not work as well on the cat's whiskers, it identified depth, while in reality there isn’t real depth there. This is as a result of a quick change in color.
The objective function for surface reconstruction can be modified to reduce noise. This can be done by introducing a smoothing term into the surface reconstruction's objective function and to the matrix.
In order to reduce the shadows weight, I multiply both sides of the equation by (I) and the results were much better.