def generateData(self): print self.frame inpt = self.getInput(0).getData() # if we are on the first frame if self.frame == 0: self.prevInpt = inpt hc = self.getInput(1).getData() # Create the features vector for t_0 storing (x,y,confidence, active) fts = numpy.ones((hc.shape[0],4), dtype=numpy.float32) fts[:,:2]=hc[:,:2] #Copy the x-y positions of features detected by hc to the features vector self.prevFeatures = fts # store it as the first entry in the features array which stores features over time self.getOutput(0).setData(fts) r = numpy.floor(self.sigmaD * 5.0/2) self.patches = numpy.zeros((hc.shape[0], (2*r+1)**2), dtype=numpy.float32) for i, (x,y, _, _) in enumerate(fts): self.patches[i,...] = inpt[y-r:y+r+1, x-r:x+r+1].flatten() else: newFeatures = numpy.copy(self.prevFeatures) Ix = self.getInput(2).getData() Iy = self.getInput(3).getData() Ixx = self.getInput(4).getData() Iyy = self.getInput(5).getData() Ixy = self.getInput(6).getData() h, w = inpt.shape velocity = numpy.array([0.0,0.0]) # intial velocity values for i, (x,y, s, a) in enumerate(newFeatures): iterations = self.iterations # throw away inactive points if a==0: continue # CALCULATE ATA pIxx = Ixx[y,x] pIyy = Iyy[y,x] pIxy = Ixy[y,x] ATA = numpy.array([[pIxx, pIxy],[pIxy, pIyy]]) #change in velocity dv = numpy.array([100.0,100.0]) g = imgutil.gaussian(self.sigmaD)[0] g = g[:,None] gg = numpy.dot(g, g.transpose()).flatten() r = g.size/2 iyy, ixx = numpy.mgrid[-r:r+1, -r:r+1] ryy, rxx = y+iyy, x+ixx patchIx = interpolation.map_coordinates(Ix, numpy.array([ryy.flatten(), rxx.flatten()])) patchIy = interpolation.map_coordinates(Iy, numpy.array([ryy.flatten(), rxx.flatten()])) # ITERATE AND CALCULATE ATb while iterations > 0 and numpy.dot(dv, dv)>self.epsilon: patch1 = interpolation.map_coordinates(inpt, numpy.array([ryy.flatten(), rxx.flatten()])) patch0 = interpolation.map_coordinates(self.prevInpt, numpy.array([(ryy-velocity[1]).flatten(), (rxx-velocity[0]).flatten()])) #imgutil.imageShow(patch0.reshape((g.size,g.size)), "p0") #imgutil.imageShow(patch1.reshape((g.size,g.size)),"p1") patchIt = patch1-patch0 pIxt = (patchIt*patchIx*gg).sum() pIyt = (patchIt*patchIy*gg).sum() # solve ATAv = -ATb ATb = -numpy.array([pIxt, pIyt]) dv = numpy.linalg.lstsq(ATA, ATb)[0] # update velocity and iterations velocity += dv iterations -= 1 #print "stopped after", (self.iterations-iterations), "with norm", numpy.dot(dv,dv)**.5 # calculate new feature positions newFeatures[i][:2]+= velocity #Compute the similarity between our new feature location and the original feature location #imgutil.imageShow(patch1.reshape(g.size,g.size), "p0") #imgutil.imageShow(self.patches[i], "p1") newFeatures[i][2] = imgutil.ncc(patch1, self.patches[i,...]) # set feature status (active or inactive) if newFeatures[i][0] > w or newFeatures[i][1] > h or newFeatures[i][0] < 0 or newFeatures[i][1] < 0: newFeatures[i][3] = 0 self.prevFeatures = newFeatures self.prevInpt = inpt self.getOutput(0).setData(newFeatures) self.frame += 1
def generateData(self):
print self.frame
inpt = self.getInput(0).getData()
# if we are on the first frame
if self.frame == 0:
self.prevInpt = inpt
hc = self.getInput(1).getData()
# Create the features vector for t_0 storing (x,y,confidence, active)
fts = numpy.ones((hc.shape[0], 4), dtype=numpy.float32)
fts[:, :
2] = hc[:, :
2] #Copy the x-y positions of features detected by hc to the features vector
self.prevFeatures = fts # store it as the first entry in the features array which stores features over time
self.getOutput(0).setData(fts)
r = numpy.floor(self.sigmaD * 5.0 / 2)
self.patches = numpy.zeros((hc.shape[0], (2 * r + 1)**2),
dtype=numpy.float32)
for i, (x, y, _, _) in enumerate(fts):
self.patches[i, ...] = inpt[y - r:y + r + 1,
x - r:x + r + 1].flatten()
else:
newFeatures = numpy.copy(self.prevFeatures)
Ix = self.getInput(2).getData()
Iy = self.getInput(3).getData()
Ixx = self.getInput(4).getData()
Iyy = self.getInput(5).getData()
Ixy = self.getInput(6).getData()
h, w = inpt.shape
velocity = numpy.array([0.0, 0.0]) # intial velocity values
for i, (x, y, s, a) in enumerate(newFeatures):
iterations = self.iterations
# throw away inactive points
if a == 0:
continue
# CALCULATE ATA
pIxx = Ixx[y, x]
pIyy = Iyy[y, x]
pIxy = Ixy[y, x]
ATA = numpy.array([[pIxx, pIxy], [pIxy, pIyy]])
#change in velocity
dv = numpy.array([100.0, 100.0])
g = imgutil.gaussian(self.sigmaD)[0]
g = g[:, None]
gg = numpy.dot(g, g.transpose()).flatten()
r = g.size / 2
iyy, ixx = numpy.mgrid[-r:r + 1, -r:r + 1]
ryy, rxx = y + iyy, x + ixx
patchIx = interpolation.map_coordinates(
Ix, numpy.array([ryy.flatten(),
rxx.flatten()]))
patchIy = interpolation.map_coordinates(
Iy, numpy.array([ryy.flatten(),
rxx.flatten()]))
# ITERATE AND CALCULATE ATb
while iterations > 0 and numpy.dot(dv, dv) > self.epsilon:
patch1 = interpolation.map_coordinates(
inpt, numpy.array([ryy.flatten(),
rxx.flatten()]))
patch0 = interpolation.map_coordinates(
self.prevInpt,
numpy.array([(ryy - velocity[1]).flatten(),
(rxx - velocity[0]).flatten()]))
#imgutil.imageShow(patch0.reshape((g.size,g.size)), "p0")
#imgutil.imageShow(patch1.reshape((g.size,g.size)),"p1")
patchIt = patch1 - patch0
pIxt = (patchIt * patchIx * gg).sum()
pIyt = (patchIt * patchIy * gg).sum()
# solve ATAv = -ATb
ATb = -numpy.array([pIxt, pIyt])
dv = numpy.linalg.lstsq(ATA, ATb)[0]
# update velocity and iterations
velocity += dv
iterations -= 1
#print "stopped after", (self.iterations-iterations), "with norm", numpy.dot(dv,dv)**.5
# calculate new feature positions
newFeatures[i][:2] += velocity
#Compute the similarity between our new feature location and the original feature location
#imgutil.imageShow(patch1.reshape(g.size,g.size), "p0")
#imgutil.imageShow(self.patches[i], "p1")
newFeatures[i][2] = imgutil.ncc(patch1, self.patches[i, ...])
# set feature status (active or inactive)
if newFeatures[i][0] > w or newFeatures[i][
1] > h or newFeatures[i][0] < 0 or newFeatures[i][
1] < 0:
newFeatures[i][3] = 0
self.prevFeatures = newFeatures
self.prevInpt = inpt
self.getOutput(0).setData(newFeatures)
self.frame += 1
def generateData(self):
print "Frame: %d" % (self.frame_number)
# if on the first frame
if self.frame_number == 0:
img = self.getInput(0).getData()
self.last_img = img
harris_corners = self.getInput(1).getData()
#NumFeaturesx4 array containing the position, strength & active flag
features = numpy.ones((harris_corners.shape[0],4), dtype=numpy.float32)
features[:,:2]=harris_corners[:,:2] #->initialize features to harris corners
self.last_features = features #
self.getOutput(0).setData(features)
self.feature_list.append(features)
#make a bank of patches
r = numpy.floor(self.sigma_d * 5.0/2)
self.patches = numpy.zeros((harris_corners.shape[0], (2*r+1)**2), dtype=numpy.float32)
for i, (x,y, _, _) in enumerate(features):
self.patches[i,...] = img[y-r:y+r+1, x-r:x+r+1].flatten()
else:
features = numpy.copy(self.last_features)
img = self.getInput(0).getData()
Ixx, Iyy, Ixy = self.getInput(2).getData()
Ix, Iy = self.getInput(3).getData()
#initialize features to be stationary
v = numpy.array([0.0, 0.0])
#loop through all features
for i in range(features.shape[0]):
#skip if not active
if features[i][3] != 0:
x,y = features[i][:2]
#compute A^T*A
A = numpy.array([[Ixx[y,x],Ixy[y,x]],
[Ixy[y,x],Iyy[y,x]]])
#velocity difference
delta_v = numpy.array([100.0,100.0])
g = imgutil.gaussian(self.sigma_d)[0]
g = g[:,None]
gg = numpy.dot(g, g.transpose()).flatten()
r = g.size/2
iyy, ixx = numpy.mgrid[-r:r+1, -r:r+1]
ryy = y+iyy
rxx = x+ixx
patchIx = interpolation.map_coordinates(Ix, numpy.array([ryy.flatten(), rxx.flatten()]))
patchIy = interpolation.map_coordinates(Iy, numpy.array([ryy.flatten(), rxx.flatten()]))
#loop through to calculate velocity
iterations = 10
eps = float('1.0e-3')**2
while iterations > 0 and numpy.dot(delta_v, delta_v)> eps:
curr_patch = interpolation.map_coordinates(img, numpy.array([ryy.flatten(), rxx.flatten()]))
prev_patch = interpolation.map_coordinates(self.last_img,
numpy.array([(ryy-v[1]).flatten(), (rxx-v[0]).flatten()]))
#find temporal derivative
patch_tderiv = curr_patch-prev_patch
pIxt = (patch_tderiv*patchIx*gg).sum()
pIyt = (patch_tderiv*patchIy*gg).sum()
#ATA = -ATb
ATb = -numpy.array([pIxt, pIyt])
delta_v = numpy.linalg.lstsq(A, ATb)[0]
#update velocities
v += delta_v
iterations -= 1
# find positions of new features
features[i][:2]+= v
# weight for similarity to initial feature patch
features[i][2] = imgutil.ncc(curr_patch, self.patches[i,...])
# make inactive if out of frame
h,w = img.shape
if features[i][0] > w or features[i][1] > h or features[i][0] < 0 or features[i][1] < 0:
features[i][3] = 0
self.last_features = features
self.last_img = img
self.getOutput(0).setData(features)
self.feature_list.append(features)
self.frame_number += 1
def generateData(self):
input = self.getInput(0).getData()
mask = self.getInput(1).getData()
r = self.r
#if there is no histogram for the initial object to be tracked, grab one
if self.hist == None:
self.hist = self.make_histogram(input, self.x, self.r)
x,y = self.x[0][:]
self.feature_patch = input[x-r:x+r,y-r:y+r].flatten()
self.getOutput().setData(self.pos)
else:
#perturb particles, clip for size, make histograms for each particle
self.x += numpy.random.uniform(-self.stepsize, self.stepsize, self.x.shape)
#clip to exclude the wall
self.x = self.x.clip(numpy.array([0,103]), numpy.array(input.shape)-1).astype(int)
# List of (x,y) positions fish is present (already excludes wall)
nonzero_positions = numpy.argwhere(mask != 0)
# Ensure that all particles are on the fish
# If a particle is off the fish, randomly choose one on its body
for i in range(self.x.shape[0]):
y,x = self.x[i,:]
if mask[y,x] == 0.0:
self.x[i,:] = random.choice(nonzero_positions)
#offset = numpy.random.randint(-self.stepsize, self.stepsize, (2))
#temp = numpy.array([y,x]).clip(numpy.array([0,103]), numpy.array(input.shape)-1)
#y,x = temp
new_hist = self.make_histogram(input, self.x, self.stepsize)
#calculate weights (as battacharyya distances)
w = self.get_weights(new_hist, self.hist)
w /= numpy.sum(w) #normalize
if self.best:
#picks the location of the best particle
new_pos = self.x[numpy.argmax(w),:]
else:
#sums the weighted particle positions
new_pos = numpy.sum(self.x.T*w, axis = 1)
self.getOutput(0).setData(new_pos)
#self.setOutput(True,1)
x, y = new_pos
self.new_patch = input[x-r:x+r,y-r:y+r].flatten()
ncc = imgutil.ncc(self.feature_patch, self.new_patch)
print "NCC = %d" % (ncc)
#if ncc value < threshold, eye is not in frame
if ncc < self.threshold:
#self.setOutput(1, False)
pass
self.x = numpy.ones((self.n,2), int) * new_pos
self.pos = new_pos
if 1./sum(w**2) < self.n/2.:
self.x = self.x[self.resample(w),:]