def srasm2(enhimgtf, edgeimgs, marks, orient, k, m, modes, maxiter):
lms,_ = lm.procrustes(marks)
mean, eigenvecs, eigenvals, lm_reduced = lm.pca_reduce(lms, modes)
stdvar = np.std(lm_reduced, axis=1)
croppedmarks = np.array([])
for i in range(13):
croppedmarks = np.append(croppedmarks,lm.transform_shape(np.reshape(marks[:,i],(marks[:,i].size,1)),-1150,-500,1,0))
for i in range(13):
croppedmarks = np.append(croppedmarks,lm.transform_shape(np.reshape(marks[:,i + 13],(marks[:,i + 13].size,1)),-1173,-500,1,0))
croppedmarks = np.reshape(croppedmarks,(croppedmarks.size / 26,26),'F')
smallermarks = np.array([])
for i in range(26):
smallermarks = np.append(smallermarks,lm.transform_shape(np.reshape(croppedmarks[:,i],(croppedmarks[:,i].size,1)),0,0,1.0 / 2,0))
smallermarks = np.reshape(smallermarks,(smallermarks.size / 26,26),'F')
itx, ity, isc, itheta = init.get_initial_transformation(enhimgtf,mean,orient)
b = np.zeros((modes,1))
imgtf = pp.apply_sobel(enhimgtf)
imgtf = cv2.pyrDown(imgtf)
edgies = np.array(map(lambda x: cv2.pyrDown(x),edgeimgs))
means, covs = gl.get_statistical_model_new(edgies,smallermarks,k)
b, ntx, nty, nsc, itheta = asm(imgtf, means, covs, b, itx / 2, ity / 2, isc / 2, itheta, k, m, stdvar, mean, eigenvecs, maxiter, orient)
return lm.transform_shape(lm.pca_reconstruct(b,mean,eigenvecs),ntx * 2,nty * 2,nsc * 2,itheta)
def match_image(imgind, orientation = 2, showground = True, modes = 5, k = 5, m = 10, maxiter = 50, multires = True):
img = io.get_enhanced_img(imgind)
imges = io.get_all_gradient_img(imgind)
colgradimg = io.greyscale_to_colour(io.get_img(imgind))
#colgradimg = io.greyscale_to_colour(pp.apply_sobel(img))
# Init evaluation
upper = None
upperground = None
lower = None
lowerground = None
itervec = np.array([0,0,0,0])
if orientation != 1:
marks = io.read_all_landmarks_by_orientation(0,imgind)
if multires:
upper, itervec[0:2] = mrasm(img, imges, marks,0, k, m, modes, maxiter)
else:
upper = srasm2(img, imges, marks,0, k, m, modes, maxiter)
draw.draw_contour(colgradimg,upper, thicc=1)
allmarks = io.read_all_landmarks_by_orientation(0)
upperground = None
upperground = np.reshape(allmarks[:,imgind - 1],(allmarks[:,imgind - 1].size,1))
upperground = lm.transform_shape(upperground,-1150,-500,1,0)
if showground:
draw.draw_contour(colgradimg,upperground,color=(0,255,0), thicc=1)
if orientation != 0:
marks = io.read_all_landmarks_by_orientation(1,imgind)
if multires:
lower, itervec[2:4] = mrasm(img, imges, marks,1, k, m, modes, maxiter)
else:
lower = srasm2(img, imges, marks,1, k, m, modes, maxiter)
draw.draw_contour(colgradimg,lower, thicc=1)
allmarks = io.read_all_landmarks_by_orientation(1)
lowerground = None
lowerground = np.reshape(allmarks[:,imgind - 1],(allmarks[:,imgind - 1].size,1))
lowerground = lm.transform_shape(lowerground,-1150,-500,1,0)
if showground:
draw.draw_contour(colgradimg,lowerground,color=(0,255,0), thicc=1)
#io.show_on_screen(colgradimg,1)
#cv2.imwrite("result_contour.png",colgradimg)
dumpimg = io.greyscale_to_colour(io.get_img(imgind))
# evaluate_results(upper, upperground, lower, lowerground, True, dumpimg)
cv2.imwrite(str(imgind) + "i" + str(maxiter) + '.png',dumpimg)
#io.show_on_screen(dumpimg,1)
return itervec
def mrasm(enhimgtf, edgeimgs, marks, orient, k, m, modes, maxiter, resdepth = 1):
itervec = np.array([0,0])
lms,_ = lm.procrustes(marks)
mean, eigenvecs, eigenvals, lm_reduced = lm.pca_reduce(np.copy(lms), 26)
coverage, modes_needed = lm.get_num_eigenvecs_needed(eigenvals)
mean, eigenvecs, eigenvals, lm_reduced = lm.pca_reduce(lms, modes_needed)
stdvar = np.std(lm_reduced, axis=1)
croppedmarks = np.array([])
for i in range(13):
croppedmarks = np.append(croppedmarks,lm.transform_shape(np.reshape(marks[:,i],(marks[:,i].size,1)),-1150,-500,1,0))
for i in range(13):
croppedmarks = np.append(croppedmarks,lm.transform_shape(np.reshape(marks[:,i + 13],(marks[:,i + 13].size,1)),-1173,-500,1,0))
croppedmarks = np.reshape(croppedmarks,(croppedmarks.size / 26,26),'F')
itx, ity, isc, itheta = init.get_initial_transformation(enhimgtf,mean,orient)
b = np.zeros((modes_needed,1))
print ''
print 'Starting multi-resolution fitting procedure for orientation ' + str(orient)
print '--------'
imgtf = pp.apply_sobel(enhimgtf)
for i in range(resdepth + 1):
print 'Step ' + str(i) + ':'
edgies = edgeimgs
limgtf = imgtf
j = 0
while resdepth - i > j:
limgtf = cv2.pyrDown(limgtf)
edgies = np.array(map(lambda x: cv2.pyrDown(x),edgies))
smallermarks = np.array([])
for l in range(26):
smallermarks = np.append(smallermarks,lm.transform_shape(np.reshape(croppedmarks[:,l],(croppedmarks[:,l].size,1)),0,0,1.0 / 2,0))
smallermarks = np.reshape(smallermarks,(smallermarks.size / 26,26),'F')
j += 1
means, covs = gl.get_statistical_model_new(edgies,smallermarks,k)
b, ntx, nty, nsc, itheta, itervec[i] = asm(limgtf, means, covs, b, float(itx) / math.pow(2.0,(resdepth-i)), float(ity) / math.pow(2.0,(resdepth-i)), float(isc) / math.pow(2.0,(resdepth-i)), itheta, k, m, stdvar, mean, eigenvecs, maxiter, orient)
itx = ntx * math.pow(2.0,(resdepth-i))
ity = nty * math.pow(2.0,(resdepth-i))
isc = nsc * math.pow(2.0,(resdepth-i))
return lm.transform_shape(lm.pca_reconstruct(b,mean,eigenvecs),itx,ity,isc,itheta), itervec
def asm(imgtf, means, covs, b, tx, ty, s, theta, k, m, stdvar, mean, eigenvecs, maxiter, orient):
numit = -1
for i in range(maxiter):
shapetf = lm.pca_reconstruct(b,mean,eigenvecs)
shapetf = lm.transform_shape(shapetf, tx, ty, s, theta)
approx, stop = find_new_points(imgtf, shapetf, means, covs, k, m, orient)
if stop:
numit = i+1
print ' Stopped at iteration ' + str(i) + ' out of ' + str(maxiter) + ' maximum.'
break
b, tx, ty, s, theta = match_model_to_target(approx, mean, eigenvecs)
#Check for plausible shapes
for p in range(b.size):
b[p, 0] = max(min(b[p, 0],3*stdvar[p]),-3*stdvar[p])
if i == maxiter - 1:
numit = maxiter
print ' Completed max number of iterations: ' + str(i + 1)
return b, tx, ty, s, theta, numit
def make_object_space_visible(ls):
return lms.transform_shape(ls, 350, 250, 2203, 0)