Source for anisotropic gradients calculation

[lucagrementieri]

Thank you for the useful repository! Could you provide some references about the formula about the anisotropic gradients and why we need to use such a formula?

In calculate_gradients if the normal is the vector [-1/np.sqrt(2), 0, 1/np.sqrt(2)] I would expect to see a left_gradient with value 1 because the increment of 1 pixel causes an increment in depth by one pixel, but instead the result is 0.29289322.

In general in calculate_gradients I would expect to use the cotangent of horizontal_angle_map to obtain left_gradients but instead the scaling uses 1 - np.sin(horizontal_angle_map). I tried replacing the formula with the one using the cotangent, but the results are not good at all.

Can you explain a bit what is wrong in my reasoning? Any hint about how to recover a metric depth?