def __init__(self):
super(BioID, self).__init__('BioID', 'faces/BioID/')
point_names = ['right_eye_pupil', 'left_eye_pupil', 'mouth_right_corner', 'mouth_left_corner',
'right_eyebrow_outer_end', 'right_eyebrow_inner_end', 'left_eyebrow_inner_end',
'left_eyebrow_outer_end', 'face_left', 'right_eye_outer_corner', 'right_eye_inner_corner',
'left_eye_inner_corner', 'left_eye_outer_corner', 'face_right', 'nose_tip', 'right_nostril',
'left_nostril', 'mouth_top_lip', 'mouth_bottom_lip', 'chin_center']
self.lstImages = []
self.keyPointsDict = []
img_dir = 'BioID-FaceDatabase-V1.2/'
self.lstImages = [os.path.join(img_dir, 'BioID_' + str(i).zfill(4) + '.pgm') for i in range(1520)]
points_dir = 'points_20/'
points_path = os.path.join(self.absolute_base_directory, points_dir)
points = read_points(self.lstImages,
lambda x: x[-14:-3].lower() + 'pts',
directory=points_path)
#Note: all points are present for all images in this dataset.
#If this assumption is incorrect, point names to coordinates will
#be incorrect.
for img_points in points:
self.keyPointsDict.append({})
prev_coord = None
for idx, point in enumerate(img_points):
if idx % 2 == 0:
prev_coord = point
else:
self.keyPointsDict[-1][point_names[idx//2]] = (prev_coord, point)
prev_coord = None
def run_bayesian_learning(points_filename, som_centers_filename, activity_centers_num, path_iterations_num):
data_points = hp.read_points(points_filename)
activity_vectors = np.array(find_activity_vectors(data_points))
# prototypical activities - centers of clusters of activities
# prototypical_activities - vector of data length where each element is index of closest prototypical activity
activity_indices, prototypical_activities = kmeans.kmeans(activity_centers_num, activity_vectors)
prototypical_states = hp.read_points(som_centers_filename)
# e-step assigns each state to the closest prototypical state
# prototypical_states - vector of data length where each element is index of closest prototypical state
state_indices = kmeans.e_step(prototypical_states, data_points)
probability_matrix = find_probability_matrix(state_indices, activity_indices, prototypical_states,
prototypical_activities)
path = np.array(
generate_path(probability_matrix, prototypical_states, prototypical_activities, state_indices,
path_iterations_num))
return path
def run_bayesian_learning(points_filename, som_centers_filename,
activity_centers_num, path_iterations_num):
data_points = hp.read_points(points_filename)
activity_vectors = np.array(find_activity_vectors(data_points))
# prototypical activities - centers of clusters of activities
# prototypical_activities - vector of data length where each element is index of closest prototypical activity
activity_indices, prototypical_activities = kmeans.kmeans(
activity_centers_num, activity_vectors)
prototypical_states = hp.read_points(som_centers_filename)
# e-step assigns each state to the closest prototypical state
# prototypical_states - vector of data length where each element is index of closest prototypical state
state_indices = kmeans.e_step(prototypical_states, data_points)
probability_matrix = find_probability_matrix(state_indices,
activity_indices,
prototypical_states,
prototypical_activities)
path = np.array(
generate_path(probability_matrix, prototypical_states,
prototypical_activities, state_indices,
path_iterations_num))
return path
def __init__(self):
super(BioID, self).__init__('BioID', 'faces/BioID/')
point_names = [
'right_eye_pupil', 'left_eye_pupil', 'mouth_right_corner',
'mouth_left_corner', 'right_eyebrow_outer_end',
'right_eyebrow_inner_end', 'left_eyebrow_inner_end',
'left_eyebrow_outer_end', 'face_left', 'right_eye_outer_corner',
'right_eye_inner_corner', 'left_eye_inner_corner',
'left_eye_outer_corner', 'face_right', 'nose_tip', 'right_nostril',
'left_nostril', 'mouth_top_lip', 'mouth_bottom_lip', 'chin_center'
]
self.lstImages = []
self.keyPointsDict = []
img_dir = 'BioID-FaceDatabase-V1.2/'
self.lstImages = [
os.path.join(img_dir, 'BioID_' + str(i).zfill(4) + '.pgm')
for i in range(1520)
]
points_dir = 'points_20/'
points_path = os.path.join(self.absolute_base_directory, points_dir)
points = read_points(self.lstImages,
lambda x: x[-14:-3].lower() + 'pts',
directory=points_path)
#Note: all points are present for all images in this dataset.
#If this assumption is incorrect, point names to coordinates will
#be incorrect.
for img_points in points:
self.keyPointsDict.append({})
prev_coord = None
for idx, point in enumerate(img_points):
if idx % 2 == 0:
prev_coord = point
else:
self.keyPointsDict[-1][point_names[idx //
2]] = (prev_coord,
point)
prev_coord = None
def learn_som(input_file, centers_number, iterations_number):
som1 = SOM(centers_number)
points = hp.read_points(input_file)
xs, ys, zs = points[:, 0], points[:, 1], points[:, 2]
for iteration in range(1, iterations_number):
for i in range(0, len(xs)):
som1.update(xs[i], ys[i], zs[i], iteration, iterations_number)
# som1.som_print()
somx = []
somy = []
somz = []
for i in range(0, len(som1.list)):
somx.append(som1.list[i].x)
somy.append(som1.list[i].y)
somz.append(som1.list[i].z)
output_file = "som_centers_" + str(centers_number)
hp.write_points(np.array(somx), np.array(somy), np.array(somz), output_file)
return somx, somy, somz
def learn_som(input_file, centers_number, iterations_number):
som1 = SOM(centers_number)
points = hp.read_points(input_file)
xs, ys, zs = points[:, 0], points[:, 1], points[:, 2]
for iteration in range(1, iterations_number):
for i in range(0, len(xs)):
som1.update(xs[i], ys[i], zs[i], iteration, iterations_number)
# som1.som_print()
somx = []
somy = []
somz = []
for i in range(0, len(som1.list)):
somx.append(som1.list[i].x)
somy.append(som1.list[i].y)
somz.append(som1.list[i].z)
output_file = "som_centers_" + str(centers_number)
hp.write_points(np.array(somx), np.array(somy), np.array(somz),
output_file)
return somx, somy, somz
args = parser.parse_args()
if args.random:
# to avoid 'ValueError("sample larger than population")' when call
# random.sample().
r_end = 100
if args.random >= r_end:
r_end = args.random + 10
x_axes = random.sample(xrange(1, r_end), args.random)
y_axes = random.sample(xrange(1, r_end), args.random)
points = zip(x_axes, y_axes)
else:
from helper import read_points
points = read_points(args.file)
convex_hull = graham_scan(points)
print 'origin points:\t', points
print 'convex hull:\t', convex_hull
if args.plot:
import matplotlib.pyplot as plt
for p in points:
color = 'r' if p in convex_hull else 'b'
plt.plot(p[0], p[1], color + 'o')
h_x_axes = [ e[0] for e in convex_hull ]
h_y_axes = [ e[1] for e in convex_hull ]
import bayesian_learning as bl
import som
import helper as hp
import os.path
if __name__ == '__main__':
# STRONGLY recommended to choose q3dm1-path1.csv
point_maps = ["q3dm1-path1.csv", "q3dm1-path2.csv"]
map_index = input(
"Enter:\n0 for q3dm1-path1.csv, \n1 for q3dm1-path2.csv")
points_filename = point_maps[map_index]
print points_filename
state_centers_num = input("Enter number of som centers (reasonable: 40-100, max 1300): ")
activity_centers_num = input("Enter number of activity centers (reasonable: 40-100, max 1300): ")
som_iterations_num = input("Enter number of learning som iterations, for this some 10 - 20 is enough: ")
path_iterations_num = input("Enter number of steps to generate path (e.g 300, 500 ...): ")
som_centers_filename = "som_centers_" + str(state_centers_num)
if os.path.isfile(som_centers_filename) == False: # if there is no yet file with same number of som centers
print "Learning started"
som.learn_som(points_filename, state_centers_num, som_iterations_num) # learn som
print "Learning finished"
path = bl.run_bayesian_learning(points_filename, som_centers_filename, activity_centers_num, path_iterations_num)
data_points = hp.read_points(points_filename)
hp.plot_points(data_points[:, 0], data_points[:, 1], data_points[:, 2], "original path")
som_centers = hp.read_points(som_centers_filename)
hp.plot_points(som_centers[:, 0], som_centers[:, 1], som_centers[:, 2], "som centers")
hp.plot_points(path[:, 0], path[:, 1], path[:, 2], "generated path")
points_filename = point_maps[map_index]
print points_filename
state_centers_num = input(
"Enter number of som centers (reasonable: 40-100, max 1300): ")
activity_centers_num = input(
"Enter number of activity centers (reasonable: 40-100, max 1300): ")
som_iterations_num = input(
"Enter number of learning som iterations, for this some 10 - 20 is enough: "
)
path_iterations_num = input(
"Enter number of steps to generate path (e.g 300, 500 ...): ")
som_centers_filename = "som_centers_" + str(state_centers_num)
if os.path.isfile(
som_centers_filename
) == False: # if there is no yet file with same number of som centers
print "Learning started"
som.learn_som(points_filename, state_centers_num,
som_iterations_num) # learn som
print "Learning finished"
path = bl.run_bayesian_learning(points_filename, som_centers_filename,
activity_centers_num, path_iterations_num)
data_points = hp.read_points(points_filename)
hp.plot_points(data_points[:, 0], data_points[:, 1], data_points[:, 2],
"original path")
som_centers = hp.read_points(som_centers_filename)
hp.plot_points(som_centers[:, 0], som_centers[:, 1], som_centers[:, 2],
"som centers")
hp.plot_points(path[:, 0], path[:, 1], path[:, 2], "generated path")