def evaluateClassifier( markerFilename, bagFilenames,
rocGraphFilename = None, accuracyGraphFilename = None,
gaussianStdDev = 0.0 ):
'''Takes a list of bag files in as data for the classifier and evaluates
the classifer against a ground truth marker file. Returns the average AUC
and optimal threshold for the classifier.'''
# Load in marker buffer for evaluating classifier
markerBuffer = MarkerBuffer.loadMarkerBuffer( markerFilename )
if markerBuffer == None:
print "Error: Unable to load marker buffer -", markerFilename
# Process each bag file in turn
dataList = []
for bagFilename in bagFilenames:
print "Reading in bag file -", bagFilename
inputSequence = InputSequence( bagFilename )
if ADD_DISTRACTORS:
distractors = [
Distractor( radius=24, startPos=( 25, 35 ), endPos=( 100, 100 ), frequency=2.0 ),
Distractor( radius=24, startPos=( 200, 200 ), endPos=( 150, 50 ), frequency=0.25 ),
Distractor( radius=24, startPos=( 188, 130 ), endPos=( 168, 258 ), frequency=0.6 ),
Distractor( radius=24, startPos=( 63, 94 ), endPos=( 170, 81 ), frequency=1.5 ),
Distractor( radius=24, startPos=( 40, 287 ), endPos=( 50, 197 ), frequency=3.0 ) ]
inputSequence.addDistractorObjects( distractors )
inputSequence.calculateOpticalFlow(
OPTICAL_FLOW_BLOCK_WIDTH, OPTICAL_FLOW_BLOCK_HEIGHT,
OPTICAL_FLOW_RANGE_WIDTH, OPTICAL_FLOW_RANGE_HEIGHT,
OPTICAL_FLOW_METHOD )
#print "Resampling data"
regularisedInputSequence = RegularisedInputSequence(
inputSequence, SAMPLES_PER_SECOND )
if gaussianStdDev > 0.0:
regularisedInputSequence.smoothOpticalFlow( gaussianStdDev )
#print "Performing cross correlation"
crossCorrelatedSequence = CrossCorrelatedSequence(
regularisedInputSequence,
MAX_CORRELATION_LAG, COMBINATION_METHOD )
#print "Building ROC Curve"
rocCurve = GripperDetectorROCCurve( crossCorrelatedSequence, markerBuffer )
print "__AUC =", rocCurve.areaUnderCurve
dataList.append( WorkingData( bagFilename, inputSequence,
regularisedInputSequence, crossCorrelatedSequence, rocCurve ) )
# Average results
if len( dataList ) > 1:
curveList = [ data.rocCurve for data in dataList ]
averageROCCurve, varianceROCCurve = ROCCurve.averageROCCurves( curveList )
else:
averageROCCurve = dataList[ 0 ].rocCurve
varianceROCCurve = None
# Plot ROC Curve
figureROC = Figure( figsize=(8,6), dpi=72 )
canvasROC = FigureCanvas( figureROC )
axisROC = figureROC.add_subplot( 111 )
axisROC.plot( averageROCCurve.falsePositiveRates, averageROCCurve.truePositiveRates )
axisROC.set_xlabel( 'False Positive Rate' )
axisROC.set_ylabel( 'True Positive Rate' )
# Add error bars
if varianceROCCurve != None:
diffBetweenErrorBars = 0.025 #1.0/(NUM_ERROR_BARS)
lastFPRValue = averageROCCurve.falsePositiveRates[ 0 ]
errorFPR = []
errorTPR = []
errorErrFPR = []
errorErrTPR = []
for i in range( len( averageROCCurve.falsePositiveRates ) ):
curFPRValue = averageROCCurve.falsePositiveRates[ i ]
if abs( curFPRValue - lastFPRValue ) >= diffBetweenErrorBars:
lastFPRValue = curFPRValue
errorFPR.append( averageROCCurve.falsePositiveRates[ i ] )
errorTPR.append( averageROCCurve.truePositiveRates[ i ] )
errorErrFPR.append( math.sqrt( varianceROCCurve.falsePositiveRates[ i ] )*2.571 )
errorErrTPR.append( math.sqrt( varianceROCCurve.truePositiveRates[ i ] )*2.571 )
axisROC.errorbar( errorFPR, errorTPR,
yerr=errorErrTPR, linestyle='None' )
#xerr=errorErrFPR, yerr=errorErrTPR, linestyle='None' )
#print errorFPR
#print errorTPR
#print lastFPRValue, averageROCCurve.falsePositiveRates[ -1 ]
axisROC.set_xlim( 0.0, 1.0 )
axisROC.set_ylim( 0.0, 1.0 )
# Plot accuracy
figureAccuracy = Figure( figsize=(8,6), dpi=72 )
canvasAccuracy = FigureCanvas( figureAccuracy )
axisAccuracy = figureAccuracy.add_subplot( 111 )
thresholds = averageROCCurve.calculateThresholds()
axisAccuracy.plot( thresholds, averageROCCurve.accuracy )
#for data in dataList:
# axisAccuracy.plot( thresholds, data.rocCurve.accuracy )
# Add error bars
if varianceROCCurve != None:
diffBetweenErrorBars = 1.0/(NUM_ERROR_BARS)
lastThresholdValue = thresholds[ 0 ]
errorThreshold = []
errorAccuracy = []
errorErrAccuracy = []
for i in range( len( thresholds ) ):
curThresholdValue = thresholds[ i ]
if abs( curThresholdValue - lastThresholdValue ) >= diffBetweenErrorBars:
lastThresholdValue = curThresholdValue
errorThreshold.append( curThresholdValue )
errorAccuracy.append( averageROCCurve.accuracy[ i ] )
errorErrAccuracy.append( math.sqrt( varianceROCCurve.accuracy[ i ] )*2.571 )
axisAccuracy.errorbar( errorThreshold, errorAccuracy, yerr=errorErrAccuracy, linestyle='None' )
# Mark the threshold with the maximum accuracy
maxAccuracyThreshold = thresholds[ np.argsort( averageROCCurve.accuracy )[ -1 ] ]
axisAccuracy.axvline( x=float( maxAccuracyThreshold ), color='red' )
axisAccuracy.set_xlabel( 'Threshold' )
axisAccuracy.set_ylabel( 'Accuracy' )
# Save the graphs
if rocGraphFilename != None:
figureROC.savefig( rocGraphFilename )
if accuracyGraphFilename != None:
figureAccuracy.savefig( accuracyGraphFilename )
return averageROCCurve.areaUnderCurve, maxAccuracyThreshold
def __init__( self, bagFilename ):
self.scriptPath = os.path.dirname( __file__ )
self.cameraImagePixBuf = None
self.bagFilename = bagFilename
self.lastImageGray = None
# Read in sequence
t1 = time.time()
self.inputSequence = InputSequence( bagFilename )
distractors = [
Distractor( radius=24, startPos=( 25, 35 ), endPos=( 100, 100 ), frequency=2.0 ),
Distractor( radius=24, startPos=( 200, 200 ), endPos=( 150, 50 ), frequency=0.25 ),
Distractor( radius=24, startPos=( 188, 130 ), endPos=( 168, 258 ), frequency=0.6 ),
Distractor( radius=24, startPos=( 63, 94 ), endPos=( 170, 81 ), frequency=1.5 ),
Distractor( radius=24, startPos=( 40, 287 ), endPos=( 50, 197 ), frequency=3.0 ) ]
#self.inputSequence.addDistractorObjects( distractors )
self.inputSequence.calculateOpticalFlow(
self.OPTICAL_FLOW_BLOCK_WIDTH, self.OPTICAL_FLOW_BLOCK_HEIGHT,
self.OPTICAL_FLOW_RANGE_WIDTH, self.OPTICAL_FLOW_RANGE_HEIGHT,
self.OPTICAL_FLOW_METHOD )
t2 = time.time()
print 'Processing sequence took %0.3f ms' % ((t2-t1)*1000.0)
# Resample sequence
t1 = time.time()
self.regularisedInputSequence = RegularisedInputSequence(
self.inputSequence, self.SAMPLES_PER_SECOND )
t2 = time.time()
print 'Resampling took %0.3f ms' % ((t2-t1)*1000.0)
t1 = time.time()
self.crossCorrelatedSequence = CrossCorrelatedSequence(
self.regularisedInputSequence,
self.MAX_CORRELATION_LAG, self.COMBINATION_METHOD )
t2 = time.time()
print 'Correlation took %0.3f ms' % ((t2-t1)*1000.0)
# Detect the input signal based on the correlation in the x and y axis
self.inputSignalDetectedArray = \
self.crossCorrelatedSequence.detectInputSequence( self.CORRELATION_THRESHOLD )
# Build a histogram for the gripper
self.gripperHistogram = cv.CreateHist(
[ 256/8, 256/8, 256/8 ], cv.CV_HIST_ARRAY, [ (0,255), (0,255), (0,255) ], 1 )
firstImage = self.inputSequence.cameraImages[ 0 ]
imageRGB = cv.CreateImageHeader( ( firstImage.shape[ 1 ], firstImage.shape[ 0 ] ), cv.IPL_DEPTH_8U, 3 )
cv.SetData( imageRGB, firstImage.data, firstImage.shape[ 1 ]*3 )
r_plane = cv.CreateMat( imageRGB.height, imageRGB.width, cv.CV_8UC1 )
g_plane = cv.CreateMat( imageRGB.height, imageRGB.width, cv.CV_8UC1 )
b_plane = cv.CreateMat( imageRGB.height, imageRGB.width, cv.CV_8UC1 )
cv.Split( imageRGB, r_plane, g_plane, b_plane, None )
planes = [ r_plane, g_plane, b_plane ]
maskArray = np.zeros(shape=( imageRGB.height, imageRGB.width ), dtype=np.uint8 )
for rowIdx in range( self.inputSignalDetectedArray.shape[ 0 ] ):
for colIdx in range( self.inputSignalDetectedArray.shape[ 1 ] ):
if self.inputSignalDetectedArray[ rowIdx, colIdx ]:
rowStartIdx = rowIdx*self.OPTICAL_FLOW_BLOCK_HEIGHT
rowEndIdx = rowStartIdx + self.OPTICAL_FLOW_BLOCK_HEIGHT
colStartIdx = colIdx*self.OPTICAL_FLOW_BLOCK_WIDTH
colEndIdx = colStartIdx + self.OPTICAL_FLOW_BLOCK_WIDTH
maskArray[ rowStartIdx:rowEndIdx, colStartIdx:colEndIdx ] = 255
cv.CalcHist( [ cv.GetImage( i ) for i in planes ],
self.gripperHistogram, 0, mask=cv.fromarray( maskArray ) )
markerBuffer = MarkerBuffer.loadMarkerBuffer( self.scriptPath + self.GROUND_TRUTH_FILENAME )
if markerBuffer == None:
raise Exception( "Unable to load marker buffer" )
self.rocCurve = GripperDetectorROCCurve( self.crossCorrelatedSequence, markerBuffer )
# Create the matplotlib graph
self.figure = Figure( figsize=(8,6), dpi=72 )
self.axisX = self.figure.add_subplot( 311 )
self.axisY = self.figure.add_subplot( 312 )
self.axisROC = self.figure.add_subplot( 313 )
self.canvas = None # Wait for GUI to be created before creating canvas
self.navToolbar = None
# Setup the GUI
builder = gtk.Builder()
builder.add_from_file( self.scriptPath + "/GUI/OpticalFlowExplorer.glade" )
self.dwgCameraImage = builder.get_object( "dwgCameraImage" )
self.window = builder.get_object( "winMain" )
self.vboxMain = builder.get_object( "vboxMain" )
self.hboxWorkArea = builder.get_object( "hboxWorkArea" )
self.adjTestPointX = builder.get_object( "adjTestPointX" )
self.adjTestPointY = builder.get_object( "adjTestPointY" )
self.adjTestPointX.set_upper( self.MAX_TEST_POINT_X )
self.adjTestPointY.set_upper( self.MAX_TEST_POINT_Y )
self.sequenceControls = builder.get_object( "sequenceControls" )
self.sequenceControls.setNumFrames( len( self.inputSequence.cameraImages ) )
self.sequenceControls.setOnFrameIdxChangedCallback( self.onSequenceControlsFrameIdxChanged )
self.setFrameIdx( 0 )
self.processOpticalFlowData()
builder.connect_signals( self )
updateLoop = self.update()
gobject.idle_add( updateLoop.next )
self.window.show()
self.window.maximize()
class MainWindow:
OPTICAL_FLOW_BLOCK_WIDTH = 16
OPTICAL_FLOW_BLOCK_HEIGHT = 16
OPTICAL_FLOW_RANGE_WIDTH = 16 # Range to look outside of a block for motion
OPTICAL_FLOW_RANGE_HEIGHT = 16
OPTICAL_FLOW_METHOD = "BlockMatching"
#OPTICAL_FLOW_METHOD = "LucasKanade"
#OPTICAL_FLOW_METHOD = "HornSchunck"
COMBINATION_METHOD = "NoChange"
#GROUND_TRUTH_FILENAME = "/../../config/TopPosGripper.yaml"
#GROUND_TRUTH_FILENAME = "/../../config/ExperimentPosGripper.yaml"
#GROUND_TRUTH_FILENAME = "/../../config/OnTablePosGripper.yaml"
GROUND_TRUTH_FILENAME = "/../../config/TightBasicWave_Gripper.yaml"
CORRELATION_THRESHOLD = 0.52
MAX_TEST_POINT_X = (640 - OPTICAL_FLOW_BLOCK_WIDTH)/OPTICAL_FLOW_BLOCK_WIDTH - 1
MAX_TEST_POINT_Y = (480 - OPTICAL_FLOW_BLOCK_HEIGHT)/OPTICAL_FLOW_BLOCK_HEIGHT - 1
SAMPLES_PER_SECOND = 30.0
MAX_CORRELATION_LAG = 2.0
GRIPPER_WAVE_FREQUENCY = 1.0 # Waves per second
GRIPPER_NUM_WAVES = 3.0
GRIPPER_WAVE_AMPLITUDE = math.radians( 20.0 )
#---------------------------------------------------------------------------
def __init__( self, bagFilename ):
self.scriptPath = os.path.dirname( __file__ )
self.cameraImagePixBuf = None
self.bagFilename = bagFilename
self.lastImageGray = None
# Read in sequence
t1 = time.time()
self.inputSequence = InputSequence( bagFilename )
distractors = [
Distractor( radius=24, startPos=( 25, 35 ), endPos=( 100, 100 ), frequency=2.0 ),
Distractor( radius=24, startPos=( 200, 200 ), endPos=( 150, 50 ), frequency=0.25 ),
Distractor( radius=24, startPos=( 188, 130 ), endPos=( 168, 258 ), frequency=0.6 ),
Distractor( radius=24, startPos=( 63, 94 ), endPos=( 170, 81 ), frequency=1.5 ),
Distractor( radius=24, startPos=( 40, 287 ), endPos=( 50, 197 ), frequency=3.0 ) ]
#self.inputSequence.addDistractorObjects( distractors )
self.inputSequence.calculateOpticalFlow(
self.OPTICAL_FLOW_BLOCK_WIDTH, self.OPTICAL_FLOW_BLOCK_HEIGHT,
self.OPTICAL_FLOW_RANGE_WIDTH, self.OPTICAL_FLOW_RANGE_HEIGHT,
self.OPTICAL_FLOW_METHOD )
t2 = time.time()
print 'Processing sequence took %0.3f ms' % ((t2-t1)*1000.0)
# Resample sequence
t1 = time.time()
self.regularisedInputSequence = RegularisedInputSequence(
self.inputSequence, self.SAMPLES_PER_SECOND )
t2 = time.time()
print 'Resampling took %0.3f ms' % ((t2-t1)*1000.0)
t1 = time.time()
self.crossCorrelatedSequence = CrossCorrelatedSequence(
self.regularisedInputSequence,
self.MAX_CORRELATION_LAG, self.COMBINATION_METHOD )
t2 = time.time()
print 'Correlation took %0.3f ms' % ((t2-t1)*1000.0)
# Detect the input signal based on the correlation in the x and y axis
self.inputSignalDetectedArray = \
self.crossCorrelatedSequence.detectInputSequence( self.CORRELATION_THRESHOLD )
# Build a histogram for the gripper
self.gripperHistogram = cv.CreateHist(
[ 256/8, 256/8, 256/8 ], cv.CV_HIST_ARRAY, [ (0,255), (0,255), (0,255) ], 1 )
firstImage = self.inputSequence.cameraImages[ 0 ]
imageRGB = cv.CreateImageHeader( ( firstImage.shape[ 1 ], firstImage.shape[ 0 ] ), cv.IPL_DEPTH_8U, 3 )
cv.SetData( imageRGB, firstImage.data, firstImage.shape[ 1 ]*3 )
r_plane = cv.CreateMat( imageRGB.height, imageRGB.width, cv.CV_8UC1 )
g_plane = cv.CreateMat( imageRGB.height, imageRGB.width, cv.CV_8UC1 )
b_plane = cv.CreateMat( imageRGB.height, imageRGB.width, cv.CV_8UC1 )
cv.Split( imageRGB, r_plane, g_plane, b_plane, None )
planes = [ r_plane, g_plane, b_plane ]
maskArray = np.zeros(shape=( imageRGB.height, imageRGB.width ), dtype=np.uint8 )
for rowIdx in range( self.inputSignalDetectedArray.shape[ 0 ] ):
for colIdx in range( self.inputSignalDetectedArray.shape[ 1 ] ):
if self.inputSignalDetectedArray[ rowIdx, colIdx ]:
rowStartIdx = rowIdx*self.OPTICAL_FLOW_BLOCK_HEIGHT
rowEndIdx = rowStartIdx + self.OPTICAL_FLOW_BLOCK_HEIGHT
colStartIdx = colIdx*self.OPTICAL_FLOW_BLOCK_WIDTH
colEndIdx = colStartIdx + self.OPTICAL_FLOW_BLOCK_WIDTH
maskArray[ rowStartIdx:rowEndIdx, colStartIdx:colEndIdx ] = 255
cv.CalcHist( [ cv.GetImage( i ) for i in planes ],
self.gripperHistogram, 0, mask=cv.fromarray( maskArray ) )
markerBuffer = MarkerBuffer.loadMarkerBuffer( self.scriptPath + self.GROUND_TRUTH_FILENAME )
if markerBuffer == None:
raise Exception( "Unable to load marker buffer" )
self.rocCurve = GripperDetectorROCCurve( self.crossCorrelatedSequence, markerBuffer )
# Create the matplotlib graph
self.figure = Figure( figsize=(8,6), dpi=72 )
self.axisX = self.figure.add_subplot( 311 )
self.axisY = self.figure.add_subplot( 312 )
self.axisROC = self.figure.add_subplot( 313 )
self.canvas = None # Wait for GUI to be created before creating canvas
self.navToolbar = None
# Setup the GUI
builder = gtk.Builder()
builder.add_from_file( self.scriptPath + "/GUI/OpticalFlowExplorer.glade" )
self.dwgCameraImage = builder.get_object( "dwgCameraImage" )
self.window = builder.get_object( "winMain" )
self.vboxMain = builder.get_object( "vboxMain" )
self.hboxWorkArea = builder.get_object( "hboxWorkArea" )
self.adjTestPointX = builder.get_object( "adjTestPointX" )
self.adjTestPointY = builder.get_object( "adjTestPointY" )
self.adjTestPointX.set_upper( self.MAX_TEST_POINT_X )
self.adjTestPointY.set_upper( self.MAX_TEST_POINT_Y )
self.sequenceControls = builder.get_object( "sequenceControls" )
self.sequenceControls.setNumFrames( len( self.inputSequence.cameraImages ) )
self.sequenceControls.setOnFrameIdxChangedCallback( self.onSequenceControlsFrameIdxChanged )
self.setFrameIdx( 0 )
self.processOpticalFlowData()
builder.connect_signals( self )
updateLoop = self.update()
gobject.idle_add( updateLoop.next )
self.window.show()
self.window.maximize()
#---------------------------------------------------------------------------
def onWinMainDestroy( self, widget, data = None ):
gtk.main_quit()
#---------------------------------------------------------------------------
def main( self ):
# All PyGTK applications must have a gtk.main(). Control ends here
# and waits for an event to occur (like a key press or mouse event).
gtk.main()
#---------------------------------------------------------------------------
def processOpticalFlowData( self ):
testX = int( self.adjTestPointX.get_value() )
testY = int( self.adjTestPointY.get_value() )
regSeq = self.regularisedInputSequence
corSeq = self.crossCorrelatedSequence
# Normalise the data ready for display
normalisedServoAngleData = Utils.normaliseSequence( regSeq.regularServoAngleData )
normalisedOpticalFlowDataX = Utils.normaliseSequence( regSeq.regularOpticalFlowArrayX[ testY ][ testX ] )
normalisedOpticalFlowDataY = Utils.normaliseSequence( regSeq.regularOpticalFlowArrayY[ testY ][ testX ] )
numCorrelationChannels = len( corSeq.correlationChannels )
# Plot graphs
self.axisX.clear()
self.axisX.plot( regSeq.regularSampleTimes, normalisedServoAngleData )
self.axisX.plot( regSeq.regularSampleTimes, normalisedOpticalFlowDataX )
if numCorrelationChannels >= 1:
correlationChannel = corSeq.correlationChannels[ 0 ][ testY ][ testX ]
self.axisX.plot( regSeq.regularSampleTimes[:len(correlationChannel)], correlationChannel )
self.axisY.clear()
self.axisY.plot( regSeq.regularSampleTimes, normalisedServoAngleData )
self.axisY.plot( regSeq.regularSampleTimes, normalisedOpticalFlowDataY )
inpSeq = self.inputSequence
self.axisY.plot( inpSeq.imageTimes, Utils.normaliseSequence( inpSeq.opticalFlowArraysY[ testY ][ testX ] ) )
if numCorrelationChannels >= 2:
correlationChannel = corSeq.correlationChannels[ 1 ][ testY ][ testX ]
self.axisY.plot( regSeq.regularSampleTimes[:len(correlationChannel)], correlationChannel )
self.axisROC.clear()
self.axisROC.plot( self.rocCurve.falsePositiveRates, self.rocCurve.truePositiveRates )
#self.axisROC.plot( self.rocCurve.thresholds, self.rocCurve.sensitivity )
#self.axisROC.plot( self.rocCurve.thresholds, self.rocCurve.specificity )
self.refreshGraphDisplay()
outputFile = open( self.scriptPath + "/../../test_results/CrossCorrelation.csv", "w" )
numSamples = len( regSeq.regularSampleTimes )
correlationChannel = corSeq.correlationChannels[ 0 ][ testY ][ testX ]
numCorrelationSamples = len( correlationChannel )
print >>outputFile, "Time,Input,Output,CrossCorrelation"
for i in range( numSamples ):
if i < numCorrelationSamples:
correlationData = correlationChannel[ i ]
else:
correlationData = 0.0
print >>outputFile, "{0},{1},{2},{3}".format(
regSeq.regularSampleTimes[ i ],
normalisedServoAngleData[ i ],
normalisedOpticalFlowDataX[ i ],
correlationData )
outputFile.close()
#---------------------------------------------------------------------------
def refreshGraphDisplay( self ):
if self.canvas != None:
self.hboxWorkArea.remove( self.canvas )
self.canvas.destroy()
self.canvas = None
if self.navToolbar != None:
self.vboxMain.remove( self.navToolbar )
self.navToolbar.destroy()
self.navToolbar = None
self.canvas = FigureCanvas( self.figure ) # a gtk.DrawingArea
self.canvas.show()
self.hboxWorkArea.pack_start( self.canvas, True, True )
self.hboxWorkArea.show()
# Navigation toolbar
self.navToolbar = NavigationToolbar( self.canvas, self.window )
self.navToolbar.lastDir = '/var/tmp/'
self.vboxMain.pack_start( self.navToolbar, expand=False, fill=False )
self.navToolbar.show()
self.vboxMain.show()
#---------------------------------------------------------------------------
def setFrameIdx( self, frameIdx ):
self.frameIdx = frameIdx
# Display the frame
image = self.inputSequence.cameraImages[ frameIdx ]
imageWidth = image.shape[ 1 ]
imageHeight = image.shape[ 0 ]
imageStep = imageWidth*3
self.cameraImagePixBuf = gtk.gdk.pixbuf_new_from_data(
image.tostring(),
gtk.gdk.COLORSPACE_RGB,
False,
8,
imageWidth,
imageHeight,
imageStep )
# Track gripper
imageRGB = cv.CreateImageHeader( ( imageWidth, imageHeight ), cv.IPL_DEPTH_8U, 3 )
cv.SetData( imageRGB, image.data, imageStep )
imageRGB = cv.CloneImage( imageRGB )
r_plane = cv.CreateMat( imageRGB.height, imageRGB.width, cv.CV_8UC1 )
g_plane = cv.CreateMat( imageRGB.height, imageRGB.width, cv.CV_8UC1 )
b_plane = cv.CreateMat( imageRGB.height, imageRGB.width, cv.CV_8UC1 )
cv.Split( imageRGB, r_plane, g_plane, b_plane, None )
planes = [ r_plane, g_plane, b_plane ]
backproject = cv.CreateImage(cv.GetSize(imageRGB), 8, 1)
# Run the cam-shift
cv.CalcArrBackProject( planes, backproject, self.gripperHistogram )
#cv.Threshold( backproject, backproject, 1, 255, cv.CV_THRESH_BINARY )
cv.CvtColor( backproject, imageRGB, cv.CV_GRAY2RGB )
#self.cameraImagePixBuf = gtk.gdk.pixbuf_new_from_data(
#imageRGB.tostring(),
#gtk.gdk.COLORSPACE_RGB,
#False,
#8,
#imageRGB.width,
#imageRGB.height,
#imageRGB.width*3 )
# Resize the drawing area if necessary
if self.dwgCameraImage.get_size_request() != ( imageWidth, imageHeight ):
self.dwgCameraImage.set_size_request( imageWidth, imageHeight )
self.dwgCameraImage.queue_draw()
#---------------------------------------------------------------------------
def onTestPointAdjustmentValueChanged( self, widget ):
self.processOpticalFlowData()
self.dwgCameraImage.queue_draw()
#---------------------------------------------------------------------------
def onSequenceControlsFrameIdxChanged( self, widget ):
self.setFrameIdx( widget.frameIdx )
#---------------------------------------------------------------------------
def onDwgCameraImageButtonPressEvent( self, widget, data ):
if self.cameraImagePixBuf != None:
imgRect = self.getImageRectangleInWidget( widget,
self.cameraImagePixBuf.get_width(), self.cameraImagePixBuf.get_height() )
self.adjTestPointX.set_value( int( ( data.x - imgRect.x )/self.OPTICAL_FLOW_BLOCK_WIDTH ) )
self.adjTestPointY.set_value( int( ( data.y - imgRect.y )/self.OPTICAL_FLOW_BLOCK_HEIGHT ) )
#---------------------------------------------------------------------------
def onDwgCameraImageExposeEvent( self, widget, data = None ):
if self.cameraImagePixBuf != None:
imgRect = self.getImageRectangleInWidget( widget,
self.cameraImagePixBuf.get_width(), self.cameraImagePixBuf.get_height() )
imgOffsetX = imgRect.x
imgOffsetY = imgRect.y
# Get the total area that needs to be redrawn
imgRect = imgRect.intersect( data.area )
srcX = imgRect.x - imgOffsetX
srcY = imgRect.y - imgOffsetY
widget.window.draw_pixbuf( widget.get_style().fg_gc[ gtk.STATE_NORMAL ],
self.cameraImagePixBuf, srcX, srcY,
imgRect.x, imgRect.y, imgRect.width, imgRect.height )
#return
# Draw an overlay to show places where the input motion has been detected
if self.inputSignalDetectedArray != None:
imageData = np.frombuffer( self.cameraImagePixBuf.get_pixels(), dtype=np.uint8 )
imageData.shape = ( self.cameraImagePixBuf.get_height(),
self.cameraImagePixBuf.get_width(), 3 )
graphicsContext = widget.window.new_gc()
graphicsContext.set_rgb_fg_color( gtk.gdk.Color( 65535, 65535, 0 ) )
blockY = imgRect.y
for y in range( self.inputSignalDetectedArray.shape[ 0 ] ):
blockX = imgRect.x
for x in range( self.inputSignalDetectedArray.shape[ 1 ] ):
if self.inputSignalDetectedArray[ y, x ]:
# Get source block as NumPy array
srcX = blockX - imgRect.x
srcY = blockY - imgRect.y
srcData = imageData[ srcY:srcY+self.OPTICAL_FLOW_BLOCK_HEIGHT,
srcX:srcX+self.OPTICAL_FLOW_BLOCK_WIDTH, : ]
# Create a modified version of the block with a yellow layer
# alpha blended over the top
yellowLayer = np.ones( ( self.OPTICAL_FLOW_BLOCK_WIDTH,
self.OPTICAL_FLOW_BLOCK_HEIGHT, 3 ) )*[255.0,255.0,0.0]*0.5
modifiedData = ( srcData.astype( np.float32 )*0.5 + yellowLayer ).astype( np.uint8 )
# Blit the modified version into the widget
modifiedPixBuf = gtk.gdk.pixbuf_new_from_array(
modifiedData, gtk.gdk.COLORSPACE_RGB, 8 )
widget.window.draw_pixbuf( widget.get_style().fg_gc[ gtk.STATE_NORMAL ],
modifiedPixBuf, 0, 0, blockX, blockY,
self.OPTICAL_FLOW_BLOCK_WIDTH, self.OPTICAL_FLOW_BLOCK_HEIGHT )
blockX += self.OPTICAL_FLOW_BLOCK_WIDTH
blockY += self.OPTICAL_FLOW_BLOCK_HEIGHT
return
# Draw the optical flow if it's available
opticalFlowX = self.inputSequence.opticalFlowArraysX[ :, :, self.frameIdx ]
opticalFlowY = self.inputSequence.opticalFlowArraysY[ :, :, self.frameIdx ]
if opticalFlowX != None and opticalFlowY != None:
testX = int( self.adjTestPointX.get_value() )
testY = int( self.adjTestPointY.get_value() )
graphicsContext = widget.window.new_gc()
graphicsContext.set_rgb_fg_color( gtk.gdk.Color( 0, 65535, 0 ) )
blockCentreY = imgRect.y + self.OPTICAL_FLOW_BLOCK_HEIGHT / 2
for y in range( opticalFlowX.shape[ 0 ] ):
blockCentreX = imgRect.x + self.OPTICAL_FLOW_BLOCK_WIDTH / 2
for x in range( opticalFlowX.shape[ 1 ] ):
if testX == x and testY == y:
# Highlight the current test point
radius = 2
arcX = int( blockCentreX - radius )
arcY = int( blockCentreY - radius )
arcWidth = arcHeight = int( radius * 2 )
drawFilledArc = False
graphicsContext.set_rgb_fg_color( gtk.gdk.Color( 65535, 65535, 65535 ) )
widget.window.draw_arc( graphicsContext,
drawFilledArc, arcX, arcY, arcWidth, arcHeight, 0, 360 * 64 )
endX = blockCentreX + opticalFlowX[ y, x ]
endY = blockCentreY + opticalFlowY[ y, x ]
if endY < blockCentreY:
# Up is red
graphicsContext.set_rgb_fg_color( gtk.gdk.Color( 65535, 0, 0 ) )
elif endY > blockCentreY:
# Down is blue
graphicsContext.set_rgb_fg_color( gtk.gdk.Color( 0, 0, 65535 ) )
else:
# Static is green
graphicsContext.set_rgb_fg_color( gtk.gdk.Color( 0, 65535, 0 ) )
widget.window.draw_line( graphicsContext,
int( blockCentreX ), int( blockCentreY ),
int( endX ), int( endY ) )
blockCentreX += self.OPTICAL_FLOW_BLOCK_WIDTH
blockCentreY += self.OPTICAL_FLOW_BLOCK_HEIGHT
#---------------------------------------------------------------------------
def getImageRectangleInWidget( self, widget, imageWidth, imageHeight ):
# Centre the image inside the widget
widgetX, widgetY, widgetWidth, widgetHeight = widget.get_allocation()
imgRect = gtk.gdk.Rectangle( 0, 0, widgetWidth, widgetHeight )
if widgetWidth > imageWidth:
imgRect.x = (widgetWidth - imageWidth) / 2
imgRect.width = imageWidth
if widgetHeight > imageHeight:
imgRect.y = (widgetHeight - imageHeight) / 2
imgRect.height = imageHeight
return imgRect
#---------------------------------------------------------------------------
def update( self ):
UPDATE_FREQUENCY = 30.0 # Updates in Hz
lastTime = time.clock()
while 1:
curTime = time.clock()
if curTime - lastTime >= 1.0 / UPDATE_FREQUENCY:
# Save the time
lastTime = curTime
yield True
yield False
def __init__(self, bagFilename):
self.scriptPath = os.path.dirname(__file__)
self.cameraImagePixBuf = None
self.bagFilename = bagFilename
self.lastImageGray = None
# Read in sequence
t1 = time.time()
self.inputSequence = InputSequence(bagFilename)
distractors = [
Distractor(radius=24,
startPos=(25, 35),
endPos=(100, 100),
frequency=2.0),
Distractor(radius=24,
startPos=(200, 200),
endPos=(150, 50),
frequency=0.25),
Distractor(radius=24,
startPos=(188, 130),
endPos=(168, 258),
frequency=0.6),
Distractor(radius=24,
startPos=(63, 94),
endPos=(170, 81),
frequency=1.5),
Distractor(radius=24,
startPos=(40, 287),
endPos=(50, 197),
frequency=3.0)
]
#self.inputSequence.addDistractorObjects( distractors )
self.inputSequence.calculateOpticalFlow(self.OPTICAL_FLOW_BLOCK_WIDTH,
self.OPTICAL_FLOW_BLOCK_HEIGHT,
self.OPTICAL_FLOW_RANGE_WIDTH,
self.OPTICAL_FLOW_RANGE_HEIGHT,
self.OPTICAL_FLOW_METHOD)
t2 = time.time()
print 'Processing sequence took %0.3f ms' % ((t2 - t1) * 1000.0)
# Resample sequence
t1 = time.time()
self.regularisedInputSequence = RegularisedInputSequence(
self.inputSequence, self.SAMPLES_PER_SECOND)
t2 = time.time()
print 'Resampling took %0.3f ms' % ((t2 - t1) * 1000.0)
t1 = time.time()
self.crossCorrelatedSequence = CrossCorrelatedSequence(
self.regularisedInputSequence, self.MAX_CORRELATION_LAG,
self.COMBINATION_METHOD)
t2 = time.time()
print 'Correlation took %0.3f ms' % ((t2 - t1) * 1000.0)
# Detect the input signal based on the correlation in the x and y axis
self.inputSignalDetectedArray = \
self.crossCorrelatedSequence.detectInputSequence( self.CORRELATION_THRESHOLD )
# Build a histogram for the gripper
self.gripperHistogram = cv.CreateHist([256 / 8, 256 / 8, 256 / 8],
cv.CV_HIST_ARRAY, [(0, 255),
(0, 255),
(0, 255)], 1)
firstImage = self.inputSequence.cameraImages[0]
imageRGB = cv.CreateImageHeader(
(firstImage.shape[1], firstImage.shape[0]), cv.IPL_DEPTH_8U, 3)
cv.SetData(imageRGB, firstImage.data, firstImage.shape[1] * 3)
r_plane = cv.CreateMat(imageRGB.height, imageRGB.width, cv.CV_8UC1)
g_plane = cv.CreateMat(imageRGB.height, imageRGB.width, cv.CV_8UC1)
b_plane = cv.CreateMat(imageRGB.height, imageRGB.width, cv.CV_8UC1)
cv.Split(imageRGB, r_plane, g_plane, b_plane, None)
planes = [r_plane, g_plane, b_plane]
maskArray = np.zeros(shape=(imageRGB.height, imageRGB.width),
dtype=np.uint8)
for rowIdx in range(self.inputSignalDetectedArray.shape[0]):
for colIdx in range(self.inputSignalDetectedArray.shape[1]):
if self.inputSignalDetectedArray[rowIdx, colIdx]:
rowStartIdx = rowIdx * self.OPTICAL_FLOW_BLOCK_HEIGHT
rowEndIdx = rowStartIdx + self.OPTICAL_FLOW_BLOCK_HEIGHT
colStartIdx = colIdx * self.OPTICAL_FLOW_BLOCK_WIDTH
colEndIdx = colStartIdx + self.OPTICAL_FLOW_BLOCK_WIDTH
maskArray[rowStartIdx:rowEndIdx,
colStartIdx:colEndIdx] = 255
cv.CalcHist([cv.GetImage(i) for i in planes],
self.gripperHistogram,
0,
mask=cv.fromarray(maskArray))
markerBuffer = MarkerBuffer.loadMarkerBuffer(
self.scriptPath + self.GROUND_TRUTH_FILENAME)
if markerBuffer == None:
raise Exception("Unable to load marker buffer")
self.rocCurve = GripperDetectorROCCurve(self.crossCorrelatedSequence,
markerBuffer)
# Create the matplotlib graph
self.figure = Figure(figsize=(8, 6), dpi=72)
self.axisX = self.figure.add_subplot(311)
self.axisY = self.figure.add_subplot(312)
self.axisROC = self.figure.add_subplot(313)
self.canvas = None # Wait for GUI to be created before creating canvas
self.navToolbar = None
# Setup the GUI
builder = gtk.Builder()
builder.add_from_file(self.scriptPath +
"/GUI/OpticalFlowExplorer.glade")
self.dwgCameraImage = builder.get_object("dwgCameraImage")
self.window = builder.get_object("winMain")
self.vboxMain = builder.get_object("vboxMain")
self.hboxWorkArea = builder.get_object("hboxWorkArea")
self.adjTestPointX = builder.get_object("adjTestPointX")
self.adjTestPointY = builder.get_object("adjTestPointY")
self.adjTestPointX.set_upper(self.MAX_TEST_POINT_X)
self.adjTestPointY.set_upper(self.MAX_TEST_POINT_Y)
self.sequenceControls = builder.get_object("sequenceControls")
self.sequenceControls.setNumFrames(len(
self.inputSequence.cameraImages))
self.sequenceControls.setOnFrameIdxChangedCallback(
self.onSequenceControlsFrameIdxChanged)
self.setFrameIdx(0)
self.processOpticalFlowData()
builder.connect_signals(self)
updateLoop = self.update()
gobject.idle_add(updateLoop.next)
self.window.show()
self.window.maximize()
class MainWindow:
OPTICAL_FLOW_BLOCK_WIDTH = 16
OPTICAL_FLOW_BLOCK_HEIGHT = 16
OPTICAL_FLOW_RANGE_WIDTH = 16 # Range to look outside of a block for motion
OPTICAL_FLOW_RANGE_HEIGHT = 16
OPTICAL_FLOW_METHOD = "BlockMatching"
#OPTICAL_FLOW_METHOD = "LucasKanade"
#OPTICAL_FLOW_METHOD = "HornSchunck"
COMBINATION_METHOD = "NoChange"
#GROUND_TRUTH_FILENAME = "/../../config/TopPosGripper.yaml"
#GROUND_TRUTH_FILENAME = "/../../config/ExperimentPosGripper.yaml"
#GROUND_TRUTH_FILENAME = "/../../config/OnTablePosGripper.yaml"
GROUND_TRUTH_FILENAME = "/../../config/TightBasicWave_Gripper.yaml"
CORRELATION_THRESHOLD = 0.52
MAX_TEST_POINT_X = (
640 - OPTICAL_FLOW_BLOCK_WIDTH) / OPTICAL_FLOW_BLOCK_WIDTH - 1
MAX_TEST_POINT_Y = (
480 - OPTICAL_FLOW_BLOCK_HEIGHT) / OPTICAL_FLOW_BLOCK_HEIGHT - 1
SAMPLES_PER_SECOND = 30.0
MAX_CORRELATION_LAG = 2.0
GRIPPER_WAVE_FREQUENCY = 1.0 # Waves per second
GRIPPER_NUM_WAVES = 3.0
GRIPPER_WAVE_AMPLITUDE = math.radians(20.0)
#---------------------------------------------------------------------------
def __init__(self, bagFilename):
self.scriptPath = os.path.dirname(__file__)
self.cameraImagePixBuf = None
self.bagFilename = bagFilename
self.lastImageGray = None
# Read in sequence
t1 = time.time()
self.inputSequence = InputSequence(bagFilename)
distractors = [
Distractor(radius=24,
startPos=(25, 35),
endPos=(100, 100),
frequency=2.0),
Distractor(radius=24,
startPos=(200, 200),
endPos=(150, 50),
frequency=0.25),
Distractor(radius=24,
startPos=(188, 130),
endPos=(168, 258),
frequency=0.6),
Distractor(radius=24,
startPos=(63, 94),
endPos=(170, 81),
frequency=1.5),
Distractor(radius=24,
startPos=(40, 287),
endPos=(50, 197),
frequency=3.0)
]
#self.inputSequence.addDistractorObjects( distractors )
self.inputSequence.calculateOpticalFlow(self.OPTICAL_FLOW_BLOCK_WIDTH,
self.OPTICAL_FLOW_BLOCK_HEIGHT,
self.OPTICAL_FLOW_RANGE_WIDTH,
self.OPTICAL_FLOW_RANGE_HEIGHT,
self.OPTICAL_FLOW_METHOD)
t2 = time.time()
print 'Processing sequence took %0.3f ms' % ((t2 - t1) * 1000.0)
# Resample sequence
t1 = time.time()
self.regularisedInputSequence = RegularisedInputSequence(
self.inputSequence, self.SAMPLES_PER_SECOND)
t2 = time.time()
print 'Resampling took %0.3f ms' % ((t2 - t1) * 1000.0)
t1 = time.time()
self.crossCorrelatedSequence = CrossCorrelatedSequence(
self.regularisedInputSequence, self.MAX_CORRELATION_LAG,
self.COMBINATION_METHOD)
t2 = time.time()
print 'Correlation took %0.3f ms' % ((t2 - t1) * 1000.0)
# Detect the input signal based on the correlation in the x and y axis
self.inputSignalDetectedArray = \
self.crossCorrelatedSequence.detectInputSequence( self.CORRELATION_THRESHOLD )
# Build a histogram for the gripper
self.gripperHistogram = cv.CreateHist([256 / 8, 256 / 8, 256 / 8],
cv.CV_HIST_ARRAY, [(0, 255),
(0, 255),
(0, 255)], 1)
firstImage = self.inputSequence.cameraImages[0]
imageRGB = cv.CreateImageHeader(
(firstImage.shape[1], firstImage.shape[0]), cv.IPL_DEPTH_8U, 3)
cv.SetData(imageRGB, firstImage.data, firstImage.shape[1] * 3)
r_plane = cv.CreateMat(imageRGB.height, imageRGB.width, cv.CV_8UC1)
g_plane = cv.CreateMat(imageRGB.height, imageRGB.width, cv.CV_8UC1)
b_plane = cv.CreateMat(imageRGB.height, imageRGB.width, cv.CV_8UC1)
cv.Split(imageRGB, r_plane, g_plane, b_plane, None)
planes = [r_plane, g_plane, b_plane]
maskArray = np.zeros(shape=(imageRGB.height, imageRGB.width),
dtype=np.uint8)
for rowIdx in range(self.inputSignalDetectedArray.shape[0]):
for colIdx in range(self.inputSignalDetectedArray.shape[1]):
if self.inputSignalDetectedArray[rowIdx, colIdx]:
rowStartIdx = rowIdx * self.OPTICAL_FLOW_BLOCK_HEIGHT
rowEndIdx = rowStartIdx + self.OPTICAL_FLOW_BLOCK_HEIGHT
colStartIdx = colIdx * self.OPTICAL_FLOW_BLOCK_WIDTH
colEndIdx = colStartIdx + self.OPTICAL_FLOW_BLOCK_WIDTH
maskArray[rowStartIdx:rowEndIdx,
colStartIdx:colEndIdx] = 255
cv.CalcHist([cv.GetImage(i) for i in planes],
self.gripperHistogram,
0,
mask=cv.fromarray(maskArray))
markerBuffer = MarkerBuffer.loadMarkerBuffer(
self.scriptPath + self.GROUND_TRUTH_FILENAME)
if markerBuffer == None:
raise Exception("Unable to load marker buffer")
self.rocCurve = GripperDetectorROCCurve(self.crossCorrelatedSequence,
markerBuffer)
# Create the matplotlib graph
self.figure = Figure(figsize=(8, 6), dpi=72)
self.axisX = self.figure.add_subplot(311)
self.axisY = self.figure.add_subplot(312)
self.axisROC = self.figure.add_subplot(313)
self.canvas = None # Wait for GUI to be created before creating canvas
self.navToolbar = None
# Setup the GUI
builder = gtk.Builder()
builder.add_from_file(self.scriptPath +
"/GUI/OpticalFlowExplorer.glade")
self.dwgCameraImage = builder.get_object("dwgCameraImage")
self.window = builder.get_object("winMain")
self.vboxMain = builder.get_object("vboxMain")
self.hboxWorkArea = builder.get_object("hboxWorkArea")
self.adjTestPointX = builder.get_object("adjTestPointX")
self.adjTestPointY = builder.get_object("adjTestPointY")
self.adjTestPointX.set_upper(self.MAX_TEST_POINT_X)
self.adjTestPointY.set_upper(self.MAX_TEST_POINT_Y)
self.sequenceControls = builder.get_object("sequenceControls")
self.sequenceControls.setNumFrames(len(
self.inputSequence.cameraImages))
self.sequenceControls.setOnFrameIdxChangedCallback(
self.onSequenceControlsFrameIdxChanged)
self.setFrameIdx(0)
self.processOpticalFlowData()
builder.connect_signals(self)
updateLoop = self.update()
gobject.idle_add(updateLoop.next)
self.window.show()
self.window.maximize()
#---------------------------------------------------------------------------
def onWinMainDestroy(self, widget, data=None):
gtk.main_quit()
#---------------------------------------------------------------------------
def main(self):
# All PyGTK applications must have a gtk.main(). Control ends here
# and waits for an event to occur (like a key press or mouse event).
gtk.main()
#---------------------------------------------------------------------------
def processOpticalFlowData(self):
testX = int(self.adjTestPointX.get_value())
testY = int(self.adjTestPointY.get_value())
regSeq = self.regularisedInputSequence
corSeq = self.crossCorrelatedSequence
# Normalise the data ready for display
normalisedServoAngleData = Utils.normaliseSequence(
regSeq.regularServoAngleData)
normalisedOpticalFlowDataX = Utils.normaliseSequence(
regSeq.regularOpticalFlowArrayX[testY][testX])
normalisedOpticalFlowDataY = Utils.normaliseSequence(
regSeq.regularOpticalFlowArrayY[testY][testX])
numCorrelationChannels = len(corSeq.correlationChannels)
# Plot graphs
self.axisX.clear()
self.axisX.plot(regSeq.regularSampleTimes, normalisedServoAngleData)
self.axisX.plot(regSeq.regularSampleTimes, normalisedOpticalFlowDataX)
if numCorrelationChannels >= 1:
correlationChannel = corSeq.correlationChannels[0][testY][testX]
self.axisX.plot(
regSeq.regularSampleTimes[:len(correlationChannel)],
correlationChannel)
self.axisY.clear()
self.axisY.plot(regSeq.regularSampleTimes, normalisedServoAngleData)
self.axisY.plot(regSeq.regularSampleTimes, normalisedOpticalFlowDataY)
inpSeq = self.inputSequence
self.axisY.plot(
inpSeq.imageTimes,
Utils.normaliseSequence(inpSeq.opticalFlowArraysY[testY][testX]))
if numCorrelationChannels >= 2:
correlationChannel = corSeq.correlationChannels[1][testY][testX]
self.axisY.plot(
regSeq.regularSampleTimes[:len(correlationChannel)],
correlationChannel)
self.axisROC.clear()
self.axisROC.plot(self.rocCurve.falsePositiveRates,
self.rocCurve.truePositiveRates)
#self.axisROC.plot( self.rocCurve.thresholds, self.rocCurve.sensitivity )
#self.axisROC.plot( self.rocCurve.thresholds, self.rocCurve.specificity )
self.refreshGraphDisplay()
outputFile = open(
self.scriptPath + "/../../test_results/CrossCorrelation.csv", "w")
numSamples = len(regSeq.regularSampleTimes)
correlationChannel = corSeq.correlationChannels[0][testY][testX]
numCorrelationSamples = len(correlationChannel)
print >> outputFile, "Time,Input,Output,CrossCorrelation"
for i in range(numSamples):
if i < numCorrelationSamples:
correlationData = correlationChannel[i]
else:
correlationData = 0.0
print >> outputFile, "{0},{1},{2},{3}".format(
regSeq.regularSampleTimes[i], normalisedServoAngleData[i],
normalisedOpticalFlowDataX[i], correlationData)
outputFile.close()
#---------------------------------------------------------------------------
def refreshGraphDisplay(self):
if self.canvas != None:
self.hboxWorkArea.remove(self.canvas)
self.canvas.destroy()
self.canvas = None
if self.navToolbar != None:
self.vboxMain.remove(self.navToolbar)
self.navToolbar.destroy()
self.navToolbar = None
self.canvas = FigureCanvas(self.figure) # a gtk.DrawingArea
self.canvas.show()
self.hboxWorkArea.pack_start(self.canvas, True, True)
self.hboxWorkArea.show()
# Navigation toolbar
self.navToolbar = NavigationToolbar(self.canvas, self.window)
self.navToolbar.lastDir = '/var/tmp/'
self.vboxMain.pack_start(self.navToolbar, expand=False, fill=False)
self.navToolbar.show()
self.vboxMain.show()
#---------------------------------------------------------------------------
def setFrameIdx(self, frameIdx):
self.frameIdx = frameIdx
# Display the frame
image = self.inputSequence.cameraImages[frameIdx]
imageWidth = image.shape[1]
imageHeight = image.shape[0]
imageStep = imageWidth * 3
self.cameraImagePixBuf = gtk.gdk.pixbuf_new_from_data(
image.tostring(), gtk.gdk.COLORSPACE_RGB, False, 8, imageWidth,
imageHeight, imageStep)
# Track gripper
imageRGB = cv.CreateImageHeader((imageWidth, imageHeight),
cv.IPL_DEPTH_8U, 3)
cv.SetData(imageRGB, image.data, imageStep)
imageRGB = cv.CloneImage(imageRGB)
r_plane = cv.CreateMat(imageRGB.height, imageRGB.width, cv.CV_8UC1)
g_plane = cv.CreateMat(imageRGB.height, imageRGB.width, cv.CV_8UC1)
b_plane = cv.CreateMat(imageRGB.height, imageRGB.width, cv.CV_8UC1)
cv.Split(imageRGB, r_plane, g_plane, b_plane, None)
planes = [r_plane, g_plane, b_plane]
backproject = cv.CreateImage(cv.GetSize(imageRGB), 8, 1)
# Run the cam-shift
cv.CalcArrBackProject(planes, backproject, self.gripperHistogram)
#cv.Threshold( backproject, backproject, 1, 255, cv.CV_THRESH_BINARY )
cv.CvtColor(backproject, imageRGB, cv.CV_GRAY2RGB)
#self.cameraImagePixBuf = gtk.gdk.pixbuf_new_from_data(
#imageRGB.tostring(),
#gtk.gdk.COLORSPACE_RGB,
#False,
#8,
#imageRGB.width,
#imageRGB.height,
#imageRGB.width*3 )
# Resize the drawing area if necessary
if self.dwgCameraImage.get_size_request() != (imageWidth, imageHeight):
self.dwgCameraImage.set_size_request(imageWidth, imageHeight)
self.dwgCameraImage.queue_draw()
#---------------------------------------------------------------------------
def onTestPointAdjustmentValueChanged(self, widget):
self.processOpticalFlowData()
self.dwgCameraImage.queue_draw()
#---------------------------------------------------------------------------
def onSequenceControlsFrameIdxChanged(self, widget):
self.setFrameIdx(widget.frameIdx)
#---------------------------------------------------------------------------
def onDwgCameraImageButtonPressEvent(self, widget, data):
if self.cameraImagePixBuf != None:
imgRect = self.getImageRectangleInWidget(
widget, self.cameraImagePixBuf.get_width(),
self.cameraImagePixBuf.get_height())
self.adjTestPointX.set_value(
int((data.x - imgRect.x) / self.OPTICAL_FLOW_BLOCK_WIDTH))
self.adjTestPointY.set_value(
int((data.y - imgRect.y) / self.OPTICAL_FLOW_BLOCK_HEIGHT))
#---------------------------------------------------------------------------
def onDwgCameraImageExposeEvent(self, widget, data=None):
if self.cameraImagePixBuf != None:
imgRect = self.getImageRectangleInWidget(
widget, self.cameraImagePixBuf.get_width(),
self.cameraImagePixBuf.get_height())
imgOffsetX = imgRect.x
imgOffsetY = imgRect.y
# Get the total area that needs to be redrawn
imgRect = imgRect.intersect(data.area)
srcX = imgRect.x - imgOffsetX
srcY = imgRect.y - imgOffsetY
widget.window.draw_pixbuf(
widget.get_style().fg_gc[gtk.STATE_NORMAL],
self.cameraImagePixBuf, srcX, srcY, imgRect.x, imgRect.y,
imgRect.width, imgRect.height)
#return
# Draw an overlay to show places where the input motion has been detected
if self.inputSignalDetectedArray != None:
imageData = np.frombuffer(self.cameraImagePixBuf.get_pixels(),
dtype=np.uint8)
imageData.shape = (self.cameraImagePixBuf.get_height(),
self.cameraImagePixBuf.get_width(), 3)
graphicsContext = widget.window.new_gc()
graphicsContext.set_rgb_fg_color(gtk.gdk.Color(
65535, 65535, 0))
blockY = imgRect.y
for y in range(self.inputSignalDetectedArray.shape[0]):
blockX = imgRect.x
for x in range(self.inputSignalDetectedArray.shape[1]):
if self.inputSignalDetectedArray[y, x]:
# Get source block as NumPy array
srcX = blockX - imgRect.x
srcY = blockY - imgRect.y
srcData = imageData[
srcY:srcY + self.OPTICAL_FLOW_BLOCK_HEIGHT,
srcX:srcX + self.OPTICAL_FLOW_BLOCK_WIDTH, :]
# Create a modified version of the block with a yellow layer
# alpha blended over the top
yellowLayer = np.ones(
(self.OPTICAL_FLOW_BLOCK_WIDTH,
self.OPTICAL_FLOW_BLOCK_HEIGHT,
3)) * [255.0, 255.0, 0.0] * 0.5
modifiedData = (srcData.astype(np.float32) * 0.5 +
yellowLayer).astype(np.uint8)
# Blit the modified version into the widget
modifiedPixBuf = gtk.gdk.pixbuf_new_from_array(
modifiedData, gtk.gdk.COLORSPACE_RGB, 8)
widget.window.draw_pixbuf(
widget.get_style().fg_gc[gtk.STATE_NORMAL],
modifiedPixBuf, 0, 0, blockX, blockY,
self.OPTICAL_FLOW_BLOCK_WIDTH,
self.OPTICAL_FLOW_BLOCK_HEIGHT)
blockX += self.OPTICAL_FLOW_BLOCK_WIDTH
blockY += self.OPTICAL_FLOW_BLOCK_HEIGHT
return
# Draw the optical flow if it's available
opticalFlowX = self.inputSequence.opticalFlowArraysX[:, :,
self.frameIdx]
opticalFlowY = self.inputSequence.opticalFlowArraysY[:, :,
self.frameIdx]
if opticalFlowX != None and opticalFlowY != None:
testX = int(self.adjTestPointX.get_value())
testY = int(self.adjTestPointY.get_value())
graphicsContext = widget.window.new_gc()
graphicsContext.set_rgb_fg_color(gtk.gdk.Color(0, 65535, 0))
blockCentreY = imgRect.y + self.OPTICAL_FLOW_BLOCK_HEIGHT / 2
for y in range(opticalFlowX.shape[0]):
blockCentreX = imgRect.x + self.OPTICAL_FLOW_BLOCK_WIDTH / 2
for x in range(opticalFlowX.shape[1]):
if testX == x and testY == y:
# Highlight the current test point
radius = 2
arcX = int(blockCentreX - radius)
arcY = int(blockCentreY - radius)
arcWidth = arcHeight = int(radius * 2)
drawFilledArc = False
graphicsContext.set_rgb_fg_color(
gtk.gdk.Color(65535, 65535, 65535))
widget.window.draw_arc(graphicsContext,
drawFilledArc, arcX, arcY,
arcWidth, arcHeight, 0,
360 * 64)
endX = blockCentreX + opticalFlowX[y, x]
endY = blockCentreY + opticalFlowY[y, x]
if endY < blockCentreY:
# Up is red
graphicsContext.set_rgb_fg_color(
gtk.gdk.Color(65535, 0, 0))
elif endY > blockCentreY:
# Down is blue
graphicsContext.set_rgb_fg_color(
gtk.gdk.Color(0, 0, 65535))
else:
# Static is green
graphicsContext.set_rgb_fg_color(
gtk.gdk.Color(0, 65535, 0))
widget.window.draw_line(graphicsContext,
int(blockCentreX),
int(blockCentreY), int(endX),
int(endY))
blockCentreX += self.OPTICAL_FLOW_BLOCK_WIDTH
blockCentreY += self.OPTICAL_FLOW_BLOCK_HEIGHT
#---------------------------------------------------------------------------
def getImageRectangleInWidget(self, widget, imageWidth, imageHeight):
# Centre the image inside the widget
widgetX, widgetY, widgetWidth, widgetHeight = widget.get_allocation()
imgRect = gtk.gdk.Rectangle(0, 0, widgetWidth, widgetHeight)
if widgetWidth > imageWidth:
imgRect.x = (widgetWidth - imageWidth) / 2
imgRect.width = imageWidth
if widgetHeight > imageHeight:
imgRect.y = (widgetHeight - imageHeight) / 2
imgRect.height = imageHeight
return imgRect
#---------------------------------------------------------------------------
def update(self):
UPDATE_FREQUENCY = 30.0 # Updates in Hz
lastTime = time.clock()
while 1:
curTime = time.clock()
if curTime - lastTime >= 1.0 / UPDATE_FREQUENCY:
# Save the time
lastTime = curTime
yield True
yield False
def evaluateClassifier(markerFilename,
bagFilenames,
rocGraphFilename=None,
accuracyGraphFilename=None,
gaussianStdDev=0.0):
'''Takes a list of bag files in as data for the classifier and evaluates
the classifer against a ground truth marker file. Returns the average AUC
and optimal threshold for the classifier.'''
# Load in marker buffer for evaluating classifier
markerBuffer = MarkerBuffer.loadMarkerBuffer(markerFilename)
if markerBuffer == None:
print "Error: Unable to load marker buffer -", markerFilename
# Process each bag file in turn
dataList = []
for bagFilename in bagFilenames:
print "Reading in bag file -", bagFilename
inputSequence = InputSequence(bagFilename)
if ADD_DISTRACTORS:
distractors = [
Distractor(radius=24,
startPos=(25, 35),
endPos=(100, 100),
frequency=2.0),
Distractor(radius=24,
startPos=(200, 200),
endPos=(150, 50),
frequency=0.25),
Distractor(radius=24,
startPos=(188, 130),
endPos=(168, 258),
frequency=0.6),
Distractor(radius=24,
startPos=(63, 94),
endPos=(170, 81),
frequency=1.5),
Distractor(radius=24,
startPos=(40, 287),
endPos=(50, 197),
frequency=3.0)
]
inputSequence.addDistractorObjects(distractors)
inputSequence.calculateOpticalFlow(OPTICAL_FLOW_BLOCK_WIDTH,
OPTICAL_FLOW_BLOCK_HEIGHT,
OPTICAL_FLOW_RANGE_WIDTH,
OPTICAL_FLOW_RANGE_HEIGHT,
OPTICAL_FLOW_METHOD)
#print "Resampling data"
regularisedInputSequence = RegularisedInputSequence(
inputSequence, SAMPLES_PER_SECOND)
if gaussianStdDev > 0.0:
regularisedInputSequence.smoothOpticalFlow(gaussianStdDev)
#print "Performing cross correlation"
crossCorrelatedSequence = CrossCorrelatedSequence(
regularisedInputSequence, MAX_CORRELATION_LAG, COMBINATION_METHOD)
#print "Building ROC Curve"
rocCurve = GripperDetectorROCCurve(crossCorrelatedSequence,
markerBuffer)
print "__AUC =", rocCurve.areaUnderCurve
dataList.append(
WorkingData(bagFilename, inputSequence, regularisedInputSequence,
crossCorrelatedSequence, rocCurve))
# Average results
if len(dataList) > 1:
curveList = [data.rocCurve for data in dataList]
averageROCCurve, varianceROCCurve = ROCCurve.averageROCCurves(
curveList)
else:
averageROCCurve = dataList[0].rocCurve
varianceROCCurve = None
# Plot ROC Curve
figureROC = Figure(figsize=(8, 6), dpi=72)
canvasROC = FigureCanvas(figureROC)
axisROC = figureROC.add_subplot(111)
axisROC.plot(averageROCCurve.falsePositiveRates,
averageROCCurve.truePositiveRates)
axisROC.set_xlabel('False Positive Rate')
axisROC.set_ylabel('True Positive Rate')
# Add error bars
if varianceROCCurve != None:
diffBetweenErrorBars = 0.025 #1.0/(NUM_ERROR_BARS)
lastFPRValue = averageROCCurve.falsePositiveRates[0]
errorFPR = []
errorTPR = []
errorErrFPR = []
errorErrTPR = []
for i in range(len(averageROCCurve.falsePositiveRates)):
curFPRValue = averageROCCurve.falsePositiveRates[i]
if abs(curFPRValue - lastFPRValue) >= diffBetweenErrorBars:
lastFPRValue = curFPRValue
errorFPR.append(averageROCCurve.falsePositiveRates[i])
errorTPR.append(averageROCCurve.truePositiveRates[i])
errorErrFPR.append(
math.sqrt(varianceROCCurve.falsePositiveRates[i]) * 2.571)
errorErrTPR.append(
math.sqrt(varianceROCCurve.truePositiveRates[i]) * 2.571)
axisROC.errorbar(errorFPR,
errorTPR,
yerr=errorErrTPR,
linestyle='None')
#xerr=errorErrFPR, yerr=errorErrTPR, linestyle='None' )
#print errorFPR
#print errorTPR
#print lastFPRValue, averageROCCurve.falsePositiveRates[ -1 ]
axisROC.set_xlim(0.0, 1.0)
axisROC.set_ylim(0.0, 1.0)
# Plot accuracy
figureAccuracy = Figure(figsize=(8, 6), dpi=72)
canvasAccuracy = FigureCanvas(figureAccuracy)
axisAccuracy = figureAccuracy.add_subplot(111)
thresholds = averageROCCurve.calculateThresholds()
axisAccuracy.plot(thresholds, averageROCCurve.accuracy)
#for data in dataList:
# axisAccuracy.plot( thresholds, data.rocCurve.accuracy )
# Add error bars
if varianceROCCurve != None:
diffBetweenErrorBars = 1.0 / (NUM_ERROR_BARS)
lastThresholdValue = thresholds[0]
errorThreshold = []
errorAccuracy = []
errorErrAccuracy = []
for i in range(len(thresholds)):
curThresholdValue = thresholds[i]
if abs(curThresholdValue -
lastThresholdValue) >= diffBetweenErrorBars:
lastThresholdValue = curThresholdValue
errorThreshold.append(curThresholdValue)
errorAccuracy.append(averageROCCurve.accuracy[i])
errorErrAccuracy.append(
math.sqrt(varianceROCCurve.accuracy[i]) * 2.571)
axisAccuracy.errorbar(errorThreshold,
errorAccuracy,
yerr=errorErrAccuracy,
linestyle='None')
# Mark the threshold with the maximum accuracy
maxAccuracyThreshold = thresholds[np.argsort(averageROCCurve.accuracy)[-1]]
axisAccuracy.axvline(x=float(maxAccuracyThreshold), color='red')
axisAccuracy.set_xlabel('Threshold')
axisAccuracy.set_ylabel('Accuracy')
# Save the graphs
if rocGraphFilename != None:
figureROC.savefig(rocGraphFilename)
if accuracyGraphFilename != None:
figureAccuracy.savefig(accuracyGraphFilename)
return averageROCCurve.areaUnderCurve, maxAccuracyThreshold