cross pixel approach

[graemeniedermayer]

There's a second approach that I've found for making normal maps. Namely calculated the cross product between x depth differences and y depth differences.

In practice it just seemed to create either very sharp normal maps or the blue just disappeared. This probably has to do with the depth_scale.

code:

import numpy as np
def computeNormals(img_output):
	# cross pixel approach
	x_offset = img_output[1:,:]
	y_offset = img_output[:,1:]
	delta_x = img_output[:-1,:] - x_offset
	delta_y = img_output[:,:-1] - y_offset
	
	normal = np.zeros(shape=(img_output.shape[0], img_output.shape[1], 3))
	# while in pixel space step
	depth_scale = 1
	# This is the simplified cross product [1,0,delta_y] cross [0,1,delta_x] 
	normal[:,:-1,0] = delta_y
	normal[:-1,:,1] = -delta_x
	normal[:,:,2] = depth_scale
	# mirror outer edges (is this the best way to deal with the -1 length?)
	normal[:,-1:,0] = normal[:,:1,0]
	normal[-1:,:,1] = normal[:1,:,1]
	# normalize result
	n = np.linalg.norm(normal, axis=2)	
	normal[:, :, 0] /= n
	normal[:, :, 1] /= n
	normal[:, :, 2] /= n
	return normal