Datasets:

DOI:
License:
kanayamaHideaki's picture
Add semantics, instances, layout_eval, preprocessing and modifying README.md.
324d4da
# --*-- coding:utf-8 --*--
import numpy as np
from utils.util import *
np.seterr(divide='ignore', invalid='ignore')
'''
z: depth image in 'centimetres'
missingMask: a mask
C: camera matrix
'''
def processDepthImage(z, missingMask, C):
yDirParam_angleThresh = np.array([45, 15]) # threshold to estimate the direction of the gravity
yDirParam_iter = np.array([5, 5])
yDirParam_y0 = np.array([0, 1, 0])
normalParam_patchSize = np.array([3, 10])
X, Y, Z = getPointCloudFromZ(z, C, 1)
# with open('pd.txt', 'w', encoding='utf-8') as f:
# for i in range(X.shape[0]):
# for j in range(X.shape[1]):
# f.write('{} {} {}\n'.format(str(X[i,j]), str(Y[i,j]), str(Z[i,j])))
# restore x-y-z position
pc = np.zeros([z.shape[0], z.shape[1], 3])
pc[:,:,0] = X
pc[:,:,1] = Y
pc[:,:,2] = Z
N1, b1 = computeNormalsSquareSupport(z/100, missingMask, normalParam_patchSize[0],
1, C, np.ones(z.shape))
N2, b2 = computeNormalsSquareSupport(z/100, missingMask, normalParam_patchSize[1],
1, C, np.ones(z.shape))
N = N1
# Compute the direction of gravity
yDir = getYDir(N2, yDirParam_angleThresh, yDirParam_iter, yDirParam_y0)
y0 = np.array([[0, 1, 0]]).T
R = getRMatrix(y0, yDir)
# rotate the pc and N
NRot = rotatePC(N, R.T)
pcRot = rotatePC(pc, R.T)
h = -pcRot[:,:,1]
yMin = np.percentile(h, 0)
if (yMin > -90):
yMin = -130
h = h - yMin
return pc, N, yDir, h, pcRot, NRot
'''
getPointCloudFromZ: use depth image and camera matrix to get pointcloud
Z is in 'centimetres'
C: camera matrix
s: is the factor by which Z has been upsampled
'''
def getPointCloudFromZ(Z, C, s=1):
h, w= Z.shape
xx, yy = np.meshgrid(np.array(range(w))+1, np.array(range(h))+1)
# color camera parameters
cc_rgb = C[0:2,2] * s # the first two lines of colomn-3, x0 and the y0
fc_rgb = np.diag(C[0:2,0:2]) * s # number on the diagonal line
x3 = np.multiply((xx - cc_rgb[0]), Z) / fc_rgb[0]
y3 = np.multiply((yy - cc_rgb[1]), Z) / fc_rgb[1]
z3 = Z
return x3, y3, z3
'''
Clip out a 2R+1 x 2R+1 window at each point and estimate
the normal from points within this window. In case the window
straddles more than a single superpixel, only take points in the
same superpixel as the centre pixel.
Input:
depthImage: in meters
missingMask: boolean mask of what data was missing
R: radius of clipping
sc: to upsample or not
superpixels: superpixel map to define bounadaries that should
not be straddled
'''
def computeNormalsSquareSupport(depthImage, missingMask, R, sc, cameraMatrix, superpixels):
depthImage = depthImage*100 # convert to centi metres
X, Y, Z = getPointCloudFromZ(depthImage, cameraMatrix, sc)
Xf = X
Yf = Y
Zf = Z
pc = np.zeros([depthImage.shape[0], depthImage.shape[1], 3])
pc[:,:,0] = Xf
pc[:,:,1] = Yf
pc[:,:,2] = Zf
XYZf = np.copy(pc)
# find missing value
ind = np.where(missingMask == 1)
X[ind] = np.nan
Y[ind] = np.nan
Z[ind] = np.nan
one_Z = np.expand_dims(1 / Z, axis=2)
X_Z = np.divide(X, Z)
Y_Z = np.divide(Y, Z)
one = np.copy(Z)
one[np.invert(np.isnan(one[:, :]))] = 1
ZZ = np.multiply(Z, Z)
X_ZZ = np.expand_dims(np.divide(X, ZZ), axis=2)
Y_ZZ = np.expand_dims(np.divide(Y, ZZ), axis=2)
X_Z_2 = np.expand_dims(np.multiply(X_Z, X_Z), axis=2)
XY_Z = np.expand_dims(np.multiply(X_Z, Y_Z), axis=2)
Y_Z_2 = np.expand_dims(np.multiply(Y_Z, Y_Z), axis=2)
AtARaw = np.concatenate((X_Z_2, XY_Z, np.expand_dims(X_Z, axis=2), Y_Z_2,
np.expand_dims(Y_Z, axis=2), np.expand_dims(one, axis=2)), axis=2)
AtbRaw = np.concatenate((X_ZZ, Y_ZZ, one_Z), axis=2)
# with clipping
AtA = filterItChopOff(np.concatenate((AtARaw, AtbRaw), axis=2), R, superpixels)
Atb = AtA[:, :, AtARaw.shape[2]:]
AtA = AtA[:, :, :AtARaw.shape[2]]
AtA_1, detAtA = invertIt(AtA)
N = mutiplyIt(AtA_1, Atb)
divide_fac = np.sqrt(np.sum(np.multiply(N, N), axis=2))
# with np.errstate(divide='ignore'):
b = np.divide(-detAtA, divide_fac)
for i in range(3):
N[:, :, i] = np.divide(N[:, :, i], divide_fac)
# Reorient the normals to point out from the scene.
# with np.errstate(invalid='ignore'):
SN = np.sign(N[:, :, 2])
SN[SN == 0] = 1
extend_SN = np.expand_dims(SN, axis=2)
extend_SN = np.concatenate((extend_SN, extend_SN, extend_SN), axis=2)
N = np.multiply(N, extend_SN)
b = np.multiply(b, SN)
sn = np.sign(np.sum(np.multiply(N, XYZf), axis=2))
sn[np.isnan(sn)] = 1
sn[sn == 0] = 1
extend_sn = np.expand_dims(sn, axis=2)
N = np.multiply(extend_sn, N)
b = np.multiply(b, sn)
return N, b