I have a general task of discriminating images of shaded cylinders that significantly vary in 3d orientation, and almost insignificantly in size and shape.
I have treated it as a regression problem on a dataset of these oriented cylinders, but my problem is with feature computation for each sample.
Basically, what is a good way to estimate the gradient orientation from the shading of cylinders on these images?
Having limited experience, I have trouble understanding what linear or non-linear filters I should apply, so that they would be gradient-sensitive and compute orientation discriminatively for this task.
The sample image is given here(each cylinder image in a dataset is distinct):