def testClassifier(classifierFilename, case):
assert case == "lena" or case == "car", "case parameter should be lena or car"
superPixelClassifier = pomio.unpickleObject(classifierFilename)
print "\n*Loaded classifier [ ", type(
superPixelClassifier), "] from:", classifierFilename
image = None
orientation = None
if case == "car":
print "*Loading MSRC car image::"
image = pomio.msrc_loadImages(
"/home/amb/dev/mrf/data/MSRC_ObjCategImageDatabase_v2",
['Images/7_3_s.bmp'])[0].m_img
orientation = "lower"
elif case == "lena":
print "*Loading Lena.jpg"
image = skimage.data.lena()
orientation = "upper"
print "*Predicting superpixel labels in image::"
numberSuperPixels = 400
superPixelCompactness = 10
[superPixelLabels,
superPixelsMask] = predictSuperPixelLabels(superPixelClassifier, image,
numberSuperPixels,
superPixelCompactness)
carSuperPixelLabels = getSuperPixelLabelledImage(image, superPixelsMask,
superPixelLabels)
plotSuperPixelImage(image, carSuperPixelLabels, orientation)
def test():
classifierLocation = "/home/amb/dev/mrf/classifiers/logisticRegression/superpixel/logReg_miniMSRC.pkl"
classifier = pomio.unpickleObject(classifierLocation)
carFile = "7_3_s.bmp"
msrcData = "/home/amb/dev/mrf/data/MSRC_ObjCategImageDatabase_v2"
car = pomio.msrc_loadImages(msrcData, ["Images/" + carFile])[0]
groundTruth = car.m_gt
mask = superPixels.getSuperPixels_SLIC(car.m_img, 400, 10)
spLabels = SuperPixelClassifier.predictSuperPixelLabels(
classifier, car.m_img, 400, 10, True)[0]
prediction = SuperPixelClassifier.getSuperPixelLabelledImage(
car.m_img, mask, spLabels)
# save prediction to file
pomio.writeMatToCSV(prediction,
"/home/amb/dev/mrf/eval/testPrediction1.labels")
results = evaluatePrediction(prediction, groundTruth, carFile)
print "\nINFO: Car test eval results::\n\t", results
classResults = evaluateClassPerformance(prediction, groundTruth)
print "\nINFO: Car test eval class results::\n\t", classResults
confusionResults = evaluateConfusionMatrix(prediction, groundTruth)
print "\nINFO: Car test eval confusion matrix results::\n\t", "Just sum up entries... ", np.sum(
confusionResults)
def testClassifier(classifierFilename, case):
assert case=="lena" or case=="car" , "case parameter should be lena or car"
superPixelClassifier = pomio.unpickleObject(classifierFilename)
print "\n*Loaded classifier [ " , type(superPixelClassifier) , "] from:" , classifierFilename
image = None
orientation = None
if case == "car":
print "*Loading MSRC car image::"
image = pomio.msrc_loadImages("/home/amb/dev/mrf/data/MSRC_ObjCategImageDatabase_v2", ['Images/7_3_s.bmp'] )[0].m_img
orientation = "lower"
elif case == "lena":
print "*Loading Lena.jpg"
image = skimage.data.lena()
orientation = "upper"
print "*Predicting superpixel labels in image::"
numberSuperPixels = 400
superPixelCompactness = 10
[superPixelLabels, superPixelsMask] = predictSuperPixelLabels(superPixelClassifier, image,numberSuperPixels, superPixelCompactness)
carSuperPixelLabels = getSuperPixelLabelledImage(image, superPixelsMask, superPixelLabels)
plotSuperPixelImage(image, carSuperPixelLabels, orientation)
def test():
# TODO use reference classifier
classifierName = "/home/amb/dev/mrf/classifiers/randomForest/superpixel/randyForest_superPixel_maxDepth15_0.6Data.pkl"
classifier = pomio.unpickleObject(classifierName)
carFile = "7_3_s.bmp"
msrcData = "/home/amb/dev/mrf/data/MSRC_ObjCategImageDatabase_v2"
car = pomio.msrc_loadImages(msrcData, ["Images/" + carFile])[0]
groundTruth = car.m_gt
mask = SuperPixels.getSuperPixels_SLIC(car.m_img, 400, 10)
spLabels = SuperPixelClassifier.predictSuperPixelLabels(
classifier, car.m_img, 400, 10)[0]
prediction = SuperPixelClassifier.getSuperPixelLabelledImage(
car.m_img, mask, spLabels)
# save prediction to file
pomio.writeMatToCSV(
prediction, "/home/amb/dev/eval/test/predict/testPrediction1.labels")
results = evaluatePrediction(prediction, groundTruth)
print "\nINFO: Car test eval results::\n\t", results
def getAdjProbs(name):
if name != None and len(name)>0:
print 'Loading adjacency probs...'
adjProbs = pomio.unpickleObject(name)
# This is actually a bunch of counts. Some will be zero, which is probably
# a sampling error, so let's offset with some default number of counts.
adjProbs += 10.0
# Now turn it into normalised probabilities.
# todo: hey but this is not normalised for default class probability!
adjProbs /= adjProbs.sum()
# transform
adjProbs = -np.log( adjProbs )
else:
adjProbs = None
return adjProbs
def getAdjProbs(name):
if name != None and len(name) > 0:
print 'Loading adjacency probs...'
adjProbs = pomio.unpickleObject(name)
# This is actually a bunch of counts. Some will be zero, which is probably
# a sampling error, so let's offset with some default number of counts.
adjProbs += 10.0
# Now turn it into normalised probabilities.
# todo: hey but this is not normalised for default class probability!
adjProbs /= adjProbs.sum()
# transform
adjProbs = -np.log(adjProbs)
else:
adjProbs = None
return adjProbs
def test():
# TODO use reference classifier
classifierName = "/home/amb/dev/mrf/classifiers/randomForest/superpixel/randyForest_superPixel_maxDepth15_0.6Data.pkl"
classifier = pomio.unpickleObject(classifierName)
carFile = "7_3_s.bmp"
msrcData = "/home/amb/dev/mrf/data/MSRC_ObjCategImageDatabase_v2"
car = pomio.msrc_loadImages(msrcData , [ "Images/" + carFile ] )[0]
groundTruth = car.m_gt
mask = SuperPixels.getSuperPixels_SLIC(car.m_img, 400, 10)
spLabels = SuperPixelClassifier.predictSuperPixelLabels(classifier, car.m_img,400,10)[0]
prediction = SuperPixelClassifier.getSuperPixelLabelledImage(car.m_img, mask, spLabels)
# save prediction to file
pomio.writeMatToCSV(prediction, "/home/amb/dev/eval/test/predict/testPrediction1.labels")
results = evaluatePrediction(prediction, groundTruth)
print "\nINFO: Car test eval results::\n\t" , results
parser.add_argument('--nshow', type=int, default=3,\
help='Max number of features to show at once on scatter plot grid')
parser.add_argument('--nstart', type=int, default=0,\
help='Index of feature to start at')
args = parser.parse_args()
import amntools
import matplotlib.pyplot as plt
import pomio
plt.interactive(1)
# Load the features and labels
if args.ftrs.endswith('.pkl'):
ftrs = pomio.unpickleObject(args.ftrs)
else:
ftrs = pomio.readMatFromCSV(args.ftrs)
N = ftrs.shape[0]
D = ftrs.shape[1]
print '%d feature vectors of dimensionality = %d' % (N, D)
if args.labs == None:
labs = None
else:
if args.labs.endswith('.pkl'):
labs = pomio.unpickleObject(args.labs)
else:
labs = pomio.readMatFromCSV(args.labs).astype(np.int32)
# show labels
import pomio
import slic
import superPixels
import skimage
import isprs
import amntools
imgRGB = amntools.readImage(args.imagefile)
if len(args.spfile) == 0:
numberSuperPixels = args.nbSuperPixels
superPixelCompactness = args.superPixelCompactness
# Turn image into superpixels.
spix = superPixels.computeSuperPixelGraph(
imgRGB, 'slic', [numberSuperPixels, superPixelCompactness])
elif args.spfile.endswith('.pkl'):
superPixelInput = pomio.unpickleObject(args.spfile)
spix = superPixelInput[0]
classProbs = superPixelInput[1]
colourMap = pomio.msrc_classToRGB
elif args.spfile.endswith('.mat'):
spix, classProbs = isprs.loadISPRSResultFromMatlab(args.spfile)
colourMap = isprs.colourMap
else:
assert False
# Display superpixel boundaries on image.
plt.imshow(
superPixels.generateImageWithSuperPixelBoundaries(imgRGB, spix.m_labels))
plt.show()
description='Apply superPixel classifier to an image.')
parser.add_argument('clfrFn', type=str, action='store', \
help='filename of pkl or csv superPixel classifier file')
parser.add_argument('infile', type=str, action='store', \
help='filename of input image to be classified')
parser.add_argument('--outfile', type=str, action='store', \
help='filename of output image to be classified')
parser.add_argument('--verbose', action='store_true')
parser.add_argument('--nbSuperPixels', type=int, default=400, \
help='Desired number of super pixels in SLIC over-segmentation')
parser.add_argument('--superPixelCompactness', type=float, default=10.0, \
help='Super pixel compactness parameter for SLIC')
args = parser.parse_args()
clfrFn = args.clfrFn
clfr = pomio.unpickleObject(clfrFn)
infile = args.infile
outfile = args.outfile
numberSuperPixels = args.nbSuperPixels
superPixelCompactness = args.superPixelCompactness
if args.verbose:
plt.interactive(1)
plt.figure()
pomio.showClassColours()
plt.figure()
print 'Classifying file ', args.infile
image = skimage.io.imread(args.infile)
[spClassPreds, spGraph
assert clfrType in ['logreg', 'randyforest'], \
'Unknown classifier type ' + clfrType
# Check can write these files.
if outfile != None:
f = open(outfile, 'w')
f.close()
clfr = None
labs = None
ftrs = None
# Load the features and labels
if infileFtrs.endswith('.pkl'):
ftrs = pomio.unpickleObject(infileFtrs)
else:
ftrs = pomio.readMatFromCSV(infileFtrs)
D = ftrs.shape[1]
print 'Feature dimensionality = ', D
if infileLabs.endswith('.pkl'):
labs = pomio.unpickleObject(infileLabs)
else:
labs = pomio.readMatFromCSV(infileLabs).astype(np.int32)
n = len(labs)
assert n == ftrs.shape[0], 'Error: there are %d labels and %d training examples' \
% ( n, ftrs.shape[0] )
assert np.all(np.isfinite(ftrs))
#classProbs = classProbs[:,:, np.array([1,2,5,3,4])-1]
else:
# super-pixel probs
spix, classProbsLUT = isprs.loadISPRSResultFromMatlab( args.matFn )
classLabs = spix.m_labels
# map the labels to HxWxC class probabilities matrix
classProbs = np.zeros( classLabs.shape + (classProbsLUT.shape[1],) )
for c in range( classProbsLUT.shape[1] ):
# Get a mask of matching pixels
classProbs[ :,:, c ] = classProbsLUT[ classLabs, c ]
classNames = isprs.classLabels
colourMap = isprs.colourMap
else:
print 'Computing class probabilities...'
print 'Loading classifier...'
clfr = pomio.unpickleObject(clfrFn)
ftype = 'classic'
classLabs, classProbs = classification.classifyImagePixels(imgRGB, clfr, \
ftype, True)
print 'done. result size = ', classProbs.shape
print ' classes = ', clfr.classes_
# Transform class probs to the correct sized matrix.
nbRows = imgRGB.shape[0]
nbCols = imgRGB.shape[1]
nbClasses = pomio.getNumClasses()
cpnew = np.zeros( (nbRows, nbCols, nbClasses) )
for i in range( classProbs.shape[2] ):
# stuff this set of probs to new label
cpnew[:,:,clfr.classes_[i]] = classProbs[:,:,i]
'Unknown classifier type ' + clfrType
# Check can write these files.
if outfile != None:
f=open(outfile,'w')
f.close()
clfr = None
labs = None
ftrs = None
# Load the features and labels
if infileFtrs.endswith('.pkl'):
ftrs = pomio.unpickleObject( infileFtrs )
else:
ftrs = pomio.readMatFromCSV( infileFtrs )
D = ftrs.shape[1]
print 'Feature dimensionality = ', D
if infileLabs.endswith('.pkl'):
labs = pomio.unpickleObject( infileLabs )
else:
labs = pomio.readMatFromCSV( infileLabs ).astype(np.int32)
n = len(labs)
assert n == ftrs.shape[0], 'Error: there are %d labels and %d training examples' \
% ( n, ftrs.shape[0] )
assert np.all( np.isfinite( ftrs ) )
# Class vars
K = args.K
spix = None
classProbs = None
adjProbs = None
precomputedMode = args.infile.endswith('.pkl') or args.infile.endswith('.mat')
if precomputedMode == True:
# If the input file is a pkl (not image) assume we've run necessary superpixel classifier, and input is class label probabilities
# Input is assumed to be a tuple [superpixels, classProbs]
print "Using pre-computed superpixels and class label probabilities"
if args.infile.endswith('.pkl'):
superPixelInput = pomio.unpickleObject(args.infile)
spix = superPixelInput[0]
classProbs = superPixelInput[1]
colourMap = pomio.msrc_classToRGB
else:
spix, classProbs = isprs.loadISPRSResultFromMatlab( args.infile )
colourMap = isprs.colourMap
else:
# Assume input image needs processing and classifying
print "Superpixel generation mode"
numberSuperPixels = args.nbSuperPixels
superPixelCompactness = args.superPixelCompactness
parser.add_argument('--nshow', type=int, default=3,\
help='Max number of features to show at once on scatter plot grid')
parser.add_argument('--nstart', type=int, default=0,\
help='Index of feature to start at')
args = parser.parse_args()
import amntools
import matplotlib.pyplot as plt
import pomio
plt.interactive(1)
# Load the features and labels
if args.ftrs.endswith('.pkl'):
ftrs = pomio.unpickleObject( args.ftrs )
else:
ftrs = pomio.readMatFromCSV( args.ftrs )
N = ftrs.shape[0]
D = ftrs.shape[1]
print '%d feature vectors of dimensionality = %d' % (N,D)
if args.labs == None:
labs = None
else:
if args.labs.endswith('.pkl'):
labs = pomio.unpickleObject( args.labs )
else:
labs = pomio.readMatFromCSV( args.labs ).astype(np.int32)
# Class vars
K = args.K
spix = None
classProbs = None
adjProbs = None
precomputedMode = args.infile.endswith('.pkl') or args.infile.endswith('.mat')
if precomputedMode == True:
# If the input file is a pkl (not image) assume we've run necessary superpixel classifier, and input is class label probabilities
# Input is assumed to be a tuple [superpixels, classProbs]
print "Using pre-computed superpixels and class label probabilities"
if args.infile.endswith('.pkl'):
superPixelInput = pomio.unpickleObject(args.infile)
spix = superPixelInput[0]
classProbs = superPixelInput[1]
colourMap = pomio.msrc_classToRGB
else:
spix, classProbs = isprs.loadISPRSResultFromMatlab(args.infile)
colourMap = isprs.colourMap
else:
# Assume input image needs processing and classifying
print "Superpixel generation mode"
numberSuperPixels = args.nbSuperPixels
superPixelCompactness = args.superPixelCompactness
imgRGB = amntools.readImage(args.infile)
from matplotlib import cm
import pomio
import slic
import superPixels
import skimage
import isprs
import amntools
imgRGB = amntools.readImage(args.imagefile)
if len(args.spfile) == 0:
numberSuperPixels = args.nbSuperPixels
superPixelCompactness = args.superPixelCompactness
# Turn image into superpixels.
spix = superPixels.computeSuperPixelGraph(imgRGB, "slic", [numberSuperPixels, superPixelCompactness])
elif args.spfile.endswith(".pkl"):
superPixelInput = pomio.unpickleObject(args.spfile)
spix = superPixelInput[0]
classProbs = superPixelInput[1]
colourMap = pomio.msrc_classToRGB
elif args.spfile.endswith(".mat"):
spix, classProbs = isprs.loadISPRSResultFromMatlab(args.spfile)
colourMap = isprs.colourMap
else:
assert False
# Display superpixel boundaries on image.
plt.imshow(superPixels.generateImageWithSuperPixelBoundaries(imgRGB, spix.m_labels))
plt.show()
