def add_lattice_layer(gm, pixel_unaries, pixel_regularizer=None, offset=0):
"""
Adds a lattice (pixel grid) layer to a GM where all edges share the same regularizer
offset is the variable index of the first new variable. Make this the count of all variables already added
Parameters:
- pixel_unaries - a 3D array of shape (width, height, n_labels).
- pixel_regularizer (optional) - a pairwise opengm function e.g. opengm.PottsFunction([2,2],0.0,beta)
"""
n_labels = pixel_unaries.shape[-1]
n_pixels = pixel_unaries.shape[0] * pixel_unaries.shape[1]
edges = opengm.secondOrderGridVis(pixel_unaries.shape[0],
pixel_unaries.shape[1])
unaries = pixel_unaries.reshape((n_pixels, n_labels))
gm = add_layer(gm,
unaries,
edges,
pairwise=pixel_regularizer,
offset=offset)
return gm
import numpy
import opengm
import time
dimx = 1000
dimy = 1000
numVar = dimx * dimy
numLabels = 2
numberOfStates = numpy.ones(numVar, dtype=opengm.index_type) * numLabels
vis2Order = opengm.secondOrderGridVis(dimx, dimy)
numFac = len(vis2Order)
randf = numpy.random.rand(numFac, numLabels, numLabels).astype(numpy.float64)
print randf.shape
print "numVar", numVar, "numFac", numFac
print "# METHOD A"
with opengm.Timer():
gm = opengm.graphicalModel(numberOfStates, operator="adder", reserveNumFactorsPerVariable=4)
gm.reserveFunctions(numFac, "explicit")
fids = gm.addFunctions(randf)
gm.addFactors(fids, vis2Order)
print "# METHOD B"
with opengm.Timer():
# (reserve reserveNumFactorsPerVariable does not make sense if we not "finalize" factors directely)
def denoiseModel(img,
norm=2,
weight=1.0,
truncate=None,
numLabels=256,
neighbourhood=4,
inpaintPixels=None,
randInpaitStartingPoint=False):
"""
this function is used to set up a graphical model similar to
**Denoising and inpainting problems:** from `Mrf- Benchmark <http://vision.middlebury.edu/MRF/results/ >`_
Args :
img : a grayscale image in the range [0,256)
norm : used norm for unaries and 2-order functions (default : 2)
weight : weight of 2-order functions (default : 1.0)
truncate : Truncate second order function at an given value (defaut : None)
numLabels : number of labels for each variable in the graphical model,
set this to a lower number to speed up inference (default : 255)
neighbourhood : neighbourhood for the second order functions, so far only 4 is allowed (default : 4)
inpaintPixels : a tuple of x and y coordinates where no unaries are added
randInpaitStartingPoint : use a random starting point for all pixels without unaries (default : False)
"""
shape = img.shape
if (img.ndim != 2):
raise RuntimeError("image must be gray")
if neighbourhood != 4:
raise RuntimeError(
"A neighbourhood other than 4 is not yet implemented")
# normalize and flatten image
iMin = numpy.min(img)
iMax = numpy.max(img)
imgNorm = ((img[:, :] - iMin) / (iMax - iMin)) * float(numLabels)
imgFlat = imgNorm.reshape(-1).astype(numpy.uint64)
# Set up Grapical Model:
numVar = int(img.size)
gm = opengm.gm([numLabels] * numVar, operator='adder')
gm.reserveFunctions(numLabels, 'explicit')
numberOfPairwiseFactors = shape[0] * (shape[1] -
1) + shape[1] * (shape[0] - 1)
gm.reserveFactors(numVar - len(inpaintPixels[0]) + numberOfPairwiseFactors)
# Set up unaries:
# - create a range of all possible labels
allPossiblePixelValues = numpy.arange(numLabels)
pixelValueRep = numpy.repeat(allPossiblePixelValues[:, numpy.newaxis],
numLabels, 1)
# - repeat [0,1,2,3,...,253,254,255] numVar times
labelRange = numpy.arange(numLabels, dtype=opengm.value_type)
labelRange = numpy.repeat(labelRange[numpy.newaxis, :], numLabels, 0)
unaries = numpy.abs(pixelValueRep - labelRange)**norm
# - add unaries to the graphical model
fids = gm.addFunctions(unaries.astype(opengm.value_type))
# add unary factors to graphical model
if (inpaintPixels is None):
for l in xrange(numLabels):
whereL = numpy.where(imgFlat == l)
gm.addFactors(fids[l], whereL[0].astype(opengm.index_type))
else:
# get vis of inpaint pixels
ipX = inpaintPixels[0]
ipY = inpaintPixels[1]
ipVi = ipX * shape[1] + ipY
for l in xrange(numLabels):
whereL = numpy.where(imgFlat == l)
notInInpaint = numpy.setdiff1d(whereL[0], ipVi)
gm.addFactors(fids[l], notInInpaint.astype(opengm.index_type))
# add ONE second order function
f = opengm.differenceFunction(shape=[numLabels, numLabels],
norm=2,
weight=weight)
fid = gm.addFunction(f)
vis2Order = opengm.secondOrderGridVis(shape[0], shape[1], True)
# add all second order factors
gm.addFactors(fid, vis2Order)
# create a starting point
startingPoint = imgFlat.copy()
if randInpaitStartingPoint:
startingPointRandom = numpy.random.randint(
0, numLabels, size=numVar).astype(opengm.index_type)
ipVi = inpaintPixels[0] * shape[1] + inpaintPixels[1]
for x in ipVi:
startingPoint[x] = startingPointRandom[x]
startingPoint[startingPoint == numLabels] = numLabels - 1
return gm, startingPoint.astype(opengm.index_type)
def denoiseModel(
img,
norm = 2,
weight = 1.0,
truncate = None,
numLabels = 256,
neighbourhood = 4,
inpaintPixels = None,
randInpaitStartingPoint = False
):
"""
this function is used to set up a graphical model similar to
**Denoising and inpainting problems:** from `Mrf- Benchmark <http://vision.middlebury.edu/MRF/results/ >`_
Args :
img : a grayscale image in the range [0,256)
norm : used norm for unaries and 2-order functions (default : 2)
weight : weight of 2-order functions (default : 1.0)
truncate : Truncate second order function at an given value (defaut : None)
numLabels : number of labels for each variable in the graphical model,
set this to a lower number to speed up inference (default : 255)
neighbourhood : neighbourhood for the second order functions, so far only 4 is allowed (default : 4)
inpaintPixels : a tuple of x and y coordinates where no unaries are added
randInpaitStartingPoint : use a random starting point for all pixels without unaries (default : False)
"""
shape = img.shape
if(img.ndim!=2):
raise RuntimeError("image must be gray")
if neighbourhood != 4 :
raise RuntimeError("A neighbourhood other than 4 is not yet implemented")
# normalize and flatten image
iMin = numpy.min(img)
iMax = numpy.max(img)
imgNorm = ((img[:,:]-iMin)/(iMax-iMin))*float(numLabels)
imgFlat = imgNorm.reshape(-1).astype(numpy.uint64)
# Set up Grapical Model:
numVar = int(img.size)
gm = opengm.gm([numLabels]*numVar,operator='adder')
gm.reserveFunctions(numLabels,'explicit')
numberOfPairwiseFactors=shape[0]*(shape[1]-1) + shape[1]*(shape[0]-1)
gm.reserveFactors(numVar-len(inpaintPixels[0]) + numberOfPairwiseFactors )
# Set up unaries:
# - create a range of all possible labels
allPossiblePixelValues=numpy.arange(numLabels)
pixelValueRep = numpy.repeat(allPossiblePixelValues[:,numpy.newaxis],numLabels,1)
# - repeat [0,1,2,3,...,253,254,255] numVar times
labelRange = numpy.arange(numLabels,dtype=opengm.value_type)
labelRange = numpy.repeat(labelRange[numpy.newaxis,:], numLabels, 0)
unaries = numpy.abs(pixelValueRep - labelRange)**norm
# - add unaries to the graphical model
fids=gm.addFunctions(unaries.astype(opengm.value_type))
# add unary factors to graphical model
if(inpaintPixels is None):
for l in xrange(numLabels):
whereL=numpy.where(imgFlat==l)
gm.addFactors(fids[l],whereL[0].astype(opengm.index_type))
else:
# get vis of inpaint pixels
ipX = inpaintPixels[0]
ipY = inpaintPixels[1]
ipVi = ipX*shape[1] + ipY
for l in xrange(numLabels):
whereL=numpy.where(imgFlat==l)
notInInpaint=numpy.setdiff1d(whereL[0],ipVi)
gm.addFactors(fids[l],notInInpaint.astype(opengm.index_type))
# add ONE second order function
f=opengm.differenceFunction(shape=[numLabels,numLabels],norm=2,weight=weight)
fid=gm.addFunction(f)
vis2Order=opengm.secondOrderGridVis(shape[0],shape[1],True)
# add all second order factors
gm.addFactors(fid,vis2Order)
# create a starting point
startingPoint = imgFlat.copy()
if randInpaitStartingPoint :
startingPointRandom = numpy.random.randint(0,numLabels,size=numVar).astype(opengm.index_type)
ipVi = inpaintPixels[0]*shape[1] + inpaintPixels[1]
for x in ipVi:
startingPoint[x]=startingPointRandom[x]
startingPoint[startingPoint==numLabels]=numLabels-1
return gm,startingPoint.astype(opengm.index_type)
if img.shape[2] != 3:
print "Image must be RGB"
sys.exit(0)
T = float(args[6])
beta = float(args[7])
imgFlat = img.reshape([-1, 3]).view(numpy.ndarray)
numVar = imgFlat.shape[0]
gm = opengm.gm(numpy.ones(numVar, dtype=opengm.label_type) * 2)
protoColor = numpy.array([args[3], args[4], args[5]],
dtype=opengm.value_type).reshape([3, -1])
protoColor = numpy.repeat(protoColor, numVar, axis=1).swapaxes(0, 1)
diffArray = numpy.sum(numpy.abs(imgFlat - protoColor), axis=1)
unaries = numpy.ones([numVar, 2], dtype=opengm.value_type)
unaries[:, 0] = T
unaries[:, 1] = diffArray
print diffArray
gm.addFactors(gm.addFunctions(unaries), numpy.arange(numVar))
regularizer = opengm.pottsFunction([2, 2], 0.0, beta)
gridVariableIndices = opengm.secondOrderGridVis(img.shape[0], img.shape[1])
fid = gm.addFunction(regularizer)
gm.addFactors(fid, gridVariableIndices)
opengm.hdf5.saveGraphicalModel(gm, args[2], "gm")
def segment_overlap_graph(pixel_unaries,
segment_map,
segment_unaries,
pixel_regularizer=None,
segment_regularizer=None,
inter_layer_regularizer=None):
"""
greates a graphical model comprised of two layers.
- The first layer is a pixel lattice
- The second layer is a region adjacency graph over segments
- Connections exist between pixels where they are overlapped by a segment
Parameters:
- pixel_unaries - a 3D array of shape (width, height, n_labels).
- segment_map - a 2d array of shape (width, height). Each element >= 0 is a segment id that maps the corresponding pixel to that id. -1 represents no segment and no corresponding node will be added
- segment_unaries - a 2d arry of shape (n_segments, n_labels)
- pixel_regularizer (optional) - a pairwise opengm function e.g. opengm.PottsFunction([2,2],0.0,beta) or list of opengm functions of length n_pixels
- segment_regularizer (optional) - a pairwise opengm function, same requirements as pixel_regularizer
- inter_layer_regularizer (optional) - a pairwise opengm function, same requirements as pixel_regularizer
"""
_check_valid_segment_map(segment_map)
_check_compatible_segment_map_unaries(segment_map, segment_unaries)
# calculate how many variables and factors will be required
n_pixels = pixel_unaries.shape[0] * pixel_unaries.shape[1]
n_segments = segment_unaries.shape[0]
n_variables = n_pixels + n_segments
n_labels_pixels = pixel_unaries.shape[-1]
n_labels_segments = segment_unaries.shape[-1]
# calculate the region adjacency graph for the segments
rag_edges = segment_map_to_rag_edges(segment_map)
rag_edges += n_pixels #segment indices start at n_pixels remember!
n_pixel_edges = (pixel_unaries.shape[0] - 1) * pixel_unaries.shape[1] + (
pixel_unaries.shape[1] - 1) * pixel_unaries.shape[0]
n_segment_edges = rag_edges.shape[0] #check this is right
n_inter_edges = n_pixels
n_edges = n_pixel_edges + n_segment_edges + n_inter_edges
# allocate space for the model and all its variables
gm = opengm.graphicalModel([n_labels_pixels] * n_pixels +
[n_labels_segments] * n_segments)
gm.reserveFunctions(
n_variables + 3,
'explicit') # the unary functions plus the 3 types of regularizer
gm.reserveFactors(n_variables + n_edges)
# add unary functions and factors
fids = gm.addFunctions(pixel_unaries.reshape([n_pixels, n_labels_pixels]))
gm.addFactors(fids, np.arange(n_pixels), finalize=False)
fids = gm.addFunctions(segment_unaries)
gm.addFactors(fids, n_pixels + np.arange(n_segments), finalize=False)
## add pairwise functions
# pixel lattice
if pixel_regularizer is not None:
fid = gm.addFunction(pixel_regularizer)
vis = opengm.secondOrderGridVis(pixel_unaries.shape[0],
pixel_unaries.shape[1])
gm.addFactors(fid, vis, finalize=False)
# segment rag
if segment_regularizer is not None:
fid = gm.addFunction(segment_regularizer)
gm.addFactors(fid, np.sort(rag_edges, axis=1), finalize=False)
# inter-layer
if inter_layer_regularizer is not None:
fid = gm.addFunction(inter_layer_regularizer)
vis = np.dstack([
np.arange(n_pixels).reshape(pixel_unaries.shape[:2]), segment_map
]).reshape((-1, 2))
vis = _remove_rows_with_negative(vis)
vis[:, 1] += n_pixels
gm.addFactors(fid, vis, finalize=False)
gm.finalize()
return gm
imgFlat = img.reshape([-1,3]).view(numpy.ndarray) numVar = imgFlat.shape[0] gm = opengm.gm(numpy.ones(numVar,dtype=opengm.label_type)*2) protoColor = numpy.array([args[3],args[4],args[5]],dtype=opengm.value_type).reshape([3,-1]) protoColor = numpy.repeat(protoColor,numVar,axis=1).swapaxes(0,1) diffArray = numpy.sum(numpy.abs(imgFlat - protoColor),axis=1) unaries = numpy.ones([numVar,2],dtype=opengm.value_type) unaries[:,0]=T unaries[:,1]=diffArray print diffArray gm.addFactors(gm.addFunctions(unaries),numpy.arange(numVar)) regularizer=opengm.pottsFunction([2,2],0.0,beta) gridVariableIndices=opengm.secondOrderGridVis(img.shape[0],img.shape[1]) fid=gm.addFunction(regularizer) gm.addFactors(fid,gridVariableIndices) print gm inf=opengm.inference.GraphCut(gm) inf.infer() arg=inf.arg().reshape(img.shape[0:2]) vigra.impex.writeImage(arg,args[2])
