Python implementation of depth filling from NYU Depth v2 toolbox

fill_depth_colorization.py

# Original Matlab code https://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html
#
#
# Python port of depth filling code from NYU toolbox
# Speed needs to be improved
#
# Uses 'pypardiso' solver 
#
import scipy
import skimage
import numpy as np
from pypardiso import spsolve
from PIL import Image

#
# fill_depth_colorization.m
# Preprocesses the kinect depth image using a gray scale version of the
# RGB image as a weighting for the smoothing. This code is a slight
# adaptation of Anat Levin's colorization code:
#
# See: www.cs.huji.ac.il/~yweiss/Colorization/
#
# Args:
#  imgRgb - HxWx3 matrix, the rgb image for the current frame. This must
#      be between 0 and 1.
#  imgDepth - HxW matrix, the depth image for the current frame in
#       absolute (meters) space.
#  alpha - a penalty value between 0 and 1 for the current depth values.

def fill_depth_colorization(imgRgb=None, imgDepthInput=None, alpha=1):
	imgIsNoise = imgDepthInput == 0
	maxImgAbsDepth = np.max(imgDepthInput)
	imgDepth = imgDepthInput / maxImgAbsDepth
	imgDepth[imgDepth > 1] = 1
	(H, W) = imgDepth.shape
	numPix = H * W
	indsM = np.arange(numPix).reshape((W, H)).transpose()
	knownValMask = (imgIsNoise == False).astype(int)
	grayImg = skimage.color.rgb2gray(imgRgb)
	winRad = 1
	len_ = 0
	absImgNdx = 0
	len_window = (2 * winRad + 1) ** 2
	len_zeros = numPix * len_window

	cols = np.zeros(len_zeros) - 1
	rows = np.zeros(len_zeros) - 1
	vals = np.zeros(len_zeros) - 1
	gvals = np.zeros(len_window) - 1

	for j in range(W):
		for i in range(H):
			nWin = 0
			for ii in range(max(0, i - winRad), min(i + winRad + 1, H)):
				for jj in range(max(0, j - winRad), min(j + winRad + 1, W)):
					if ii == i and jj == j:
						continue

					rows[len_] = absImgNdx
					cols[len_] = indsM[ii, jj]
					gvals[nWin] = grayImg[ii, jj]

					len_ = len_ + 1
					nWin = nWin + 1

			curVal = grayImg[i, j]
			gvals[nWin] = curVal
			c_var = np.mean((gvals[:nWin + 1] - np.mean(gvals[:nWin+ 1])) ** 2)

			csig = c_var * 0.6
			mgv = np.min((gvals[:nWin] - curVal) ** 2)
			if csig < -mgv / np.log(0.01):
				csig = -mgv / np.log(0.01)

			if csig < 2e-06:
				csig = 2e-06

			gvals[:nWin] = np.exp(-(gvals[:nWin] - curVal) ** 2 / csig)
			gvals[:nWin] = gvals[:nWin] / sum(gvals[:nWin])
			vals[len_ - nWin:len_] = -gvals[:nWin]

	  		# Now the self-reference (along the diagonal).
			rows[len_] = absImgNdx
			cols[len_] = absImgNdx
			vals[len_] = 1  # sum(gvals(1:nWin))

			len_ = len_ + 1
			absImgNdx = absImgNdx + 1

	vals = vals[:len_]
	cols = cols[:len_]
	rows = rows[:len_]
	A = scipy.sparse.csr_matrix((vals, (rows, cols)), (numPix, numPix))

	rows = np.arange(0, numPix)
	cols = np.arange(0, numPix)
	vals = (knownValMask * alpha).transpose().reshape(numPix)
	G = scipy.sparse.csr_matrix((vals, (rows, cols)), (numPix, numPix))

	A = A + G
	b = np.multiply(vals.reshape(numPix), imgDepth.flatten('F'))

	#print ('Solving system..')

	new_vals = spsolve(A, b)
	new_vals = np.reshape(new_vals, (H, W), 'F')

	#print ('Done.')

	denoisedDepthImg = new_vals * maxImgAbsDepth
    
	output = denoisedDepthImg.reshape((H, W)).astype('float32')

	output = np.multiply(output, (1-knownValMask)) + imgDepthInput
    
	return output

Yes, normalize RGB to values between 0.0 and 1.0 (from 0 to 255 range).

@ialhashim
please tell how you created 80 nodes for parallel running
its taking 1 minute per image in dgx 1
share if any codes regarding cluster creation

By chance do you know if in the original colorization algorithm, what is the b vector? It seems you are using the b vector as the noisy depth image.

replace from pypardiso import spsolve with from scipy.sparse.linalg import spsolve if your not using conda.

Hello, Does this code only run on cpu？I want to implement it with pytorch

I want to ask if there is any improvement for speed?