def quat_math(q0, q1, inv0=False, inv1=False):
""" Performs addition and subtraction between quaternions
:param q0: a vector of length 4, quaternion rotation (x,y,z,w)
:param q1: a vector of length 4, quaternion rotation (x,y,z,w)
:param inv0: if True, inverts q0
:param inv1: if True, inverts q1
Examples:
to get the total rotation from going to q0 then q1: quat_math(q0,q1,False,False)
to get the rotation from q1 to q0: quat_math(q0,q1,True,False)
"""
if not isinstance(q0, np.ndarray):
q0 = np.array(q0)
if not isinstance(q1, np.ndarray):
q1 = np.array(q1)
q0 = to_transquat(q0)
q0 = trans.unit_vector(q0)
q1 = to_transquat(q1)
q1 = trans.unit_vector(q1)
if inv0:
q0 = trans.quaternion_conjugate(q0)
if inv1:
q1 = trans.quaternion_conjugate(q1)
res = trans.quaternion_multiply(q0, q1)
return to_pyquat(res)
def __init__(self, length, axis_normal, x, y, z):
self.length = length
self.axes = []
for vector in [x, y, z]:
self.axes.append(append(tf.unit_vector(array(vector)), [1]))
self.axis_normal = append(tf.unit_vector(array(axis_normal)), [1])
self.backup = [self.axis_normal, self.axes]
def user_pan(self, dx, dy):
""" Pans the camera based on passed mouse movements.
user_pan(int, int) -> None
"""
sensetivity = self._settings["pan_sensetivity"]
focus_to_cam = self._position - self._focus
x = [self._position, self._focus]
for i in range(len(x)):
x[i] = x[i] + (-sensetivity*dy * tf.unit_vector(self._up))
x[i] = x[i] + (sensetivity*dx * tf.unit_vector(
cross(focus_to_cam, self._up)))
self._position, self._focus = x
def tops(self, tops):
if tops is None:
v1, v2, v3 = list(self.verteces)
n1 = v2.coord - v1.coord
n2 = v3.coord - v1.coord
vec = np.cross(n1, n2)
d = 84 * 0.3
top1 = Top(self.coord + d * unit_vector(vec))
top2 = Top(self.coord - d * unit_vector(vec))
top1.trimer = self
top2.trimer = self
tops = set([top1, top2])
else:
assert (len(tops) == 2)
self.__tops = tops
def qv_mult(q1, v1):
v1 = transformations.unit_vector(v1)
q2 = list(v1)
q2.append(0.0)
return transformations.quaternion_multiply(
transformations.quaternion_multiply(q1, q2),
transformations.quaternion_conjugate(q1))[:3]
def user_orbit(self, dx, dy):
""" Orbits the camera based on passed mouse movements.
user_pan(int, int) -> None
Note: h denotes horizontal and v denotes vertical
"""
sensetivity = self._settings["orbit_sensetivity"]
focus_to_cam = self._position - self._focus
dist_to_focus_h = norm(focus_to_cam[:2])
dist_to_focus = norm(focus_to_cam)
angleh = atan2(focus_to_cam[1], focus_to_cam[0])
anglev = atan2(focus_to_cam[2], dist_to_focus_h)
angleh -= sensetivity*dx
anglev -= sensetivity*dy
if anglev >= pi/2-0.1:
anglev = pi/2-0.1
if anglev <= -pi/2+0.1:
anglev = -pi/2+0.1
focus_to_cam = array(
[cos(angleh)*cos(anglev)*dist_to_focus,
sin(angleh)*cos(anglev)*dist_to_focus,
sin(anglev)*dist_to_focus]
)
self._up = tf.unit_vector(-1 * cross(focus_to_cam,
cross(focus_to_cam, array([0, 0, 1]))))
self._position = self._focus + focus_to_cam
def strafe(self, ds):
strafe_vec = vector_product([0, 1, 0], self.ref)
strafe_vec = unit_vector(strafe_vec)
x = self.pos[0] + strafe_vec[0] * ds
y = self.pos[1] + strafe_vec[1] * ds
z = self.pos[2] + strafe_vec[2] * ds
self.pos = [x, y, z]
def projectVectorsImageSBA(self, IK, image):
vec_list = []
body2ned = image.get_body2ned_sba()
cam2body = image.get_cam2body()
for uv in image.uv_list:
uvh = np.array([uv[0], uv[1], 1.0])
proj = body2ned.dot(cam2body).dot(IK).dot(uvh)
proj_norm = transformations.unit_vector(proj)
vec_list.append(proj_norm)
return vec_list
def user_zoom(self, scroll_y):
""" Zooms the camera based on passed mouse movements.
user_zoom(int) -> None
"""
sensetivity = self._settings["zoom_sensetivity"]
u_focus_to_cam = tf.unit_vector(self._position - self._focus)
self._position = self._position + \
scroll_y*sensetivity * u_focus_to_cam
if norm(self._position - self._focus) <= 1:
self._focus = self._position + -1 * u_focus_to_cam
return None
def projectVectors(self, IK, body2ned, cam2body, uv_list):
proj_list = []
for uv in uv_list:
uvh = np.array([uv[0], uv[1], 1.0])
proj = body2ned.dot(cam2body).dot(IK).dot(uvh)
proj_norm = transformations.unit_vector(proj)
proj_list.append(proj_norm)
#for uv in uv_list:
# print "uv:", uv
# uvh = np.array([uv[0], uv[1], 1.0])
# print "cam vec=", transformations.unit_vector(IR.dot(IK).dot(uvh))
return proj_list
def __init__(self, T):
if ( type(T) == str):
Tfile = T
#Load transformation
Tfis = open(Tfile, 'r')
lines = []
lines = Tfis.readlines()
self.scale = float(lines[0])
self.Ss = tf.scale_matrix(self.scale)
quat_line = lines[1].split(" ")
self.quat = tf.unit_vector(np.array([float(quat_line[3]),
float(quat_line[0]),
float(quat_line[1]),
float(quat_line[2])]))
self.Hs = tf.quaternion_matrix(self.quat)
trans_line = lines[2].split(" ")
self.Ts = np.array([float(trans_line[0]),
float(trans_line[1]),
float(trans_line[2])])
Tfis.close()
self.Rs = self.Hs.copy()[:3, :3]
self.Hs[:3, 3] = self.Ts[:3]
self.Hs = self.Ss.dot(self.Hs) # to add again
elif (type(T) == np.ndarray):
self.Hs = T
scale, shear, angles, trans, persp = tf.decompose_matrix(T)
self.quat = tf.quaternion_from_euler(angles[0], angles[1], angles[2])
self.Rs = tf.quaternion_matrix(self.quat)
self.scale =scale[0]
self.Ts = trans / self.scale
print "Loaded Ground Truth Transformation: "
print self.Hs
def add(self,
particle,
template,
adding_edge,
debug=False): # returns a candidate new vertex coord
# V4,V5,V6,V7
#
# V1-------V3
# / \ /
# / \ /
# V5------V2
#
# E1 is midpoint of V1,V2 (the adding edge)
# V3 is the vertex opposite the adding edge
# V5 is the new vertex we need to calculate
#
#V4 will be an intermediate point along the axis E1-V3 which is the appropriate stem length for the template
# (essentially making the E1V3 distance longer or shorter)
#
# The convention I used is upper case V is the Vertex object. Lower case v is the actual coordinate (a numpy vector)
assert (self.full == False)
E1 = adding_edge
(V1, V2) = adding_edge.verteces
V3 = self.opposite_vertex(adding_edge)
# # maybe extend length a bit along the V1V3 vector to get the appropriate stem length for the template
E1V3 = E1.coord - V3.coord
V1V3_dist = np.linalg.norm(V3.coord - E1.coord)
d = template.stem_length / V1V3_dist
v4_1 = E1.coord - d * E1V3
v4_2 = E1.coord + d * E1V3
dist1 = np.linalg.norm(v4_1 - V3.coord)
dist2 = np.linalg.norm(v4_2 - V3.coord)
if dist1 < dist2: # pick the value closer to V3 to become V4
v4 = v4_1
else:
v4 = v4_2
# set origin of rotation to E1
e1 = E1.coord
v4 = v4 - e1
# get two new points with consistent angle for first rotation
axis = V1.coord - V2.coord
axis = unit_vector(axis)
a1 = template.angle_radians
a2 = -1 * (template.angle_radians)
v5_1 = self.get_new_point(e1, v4, a1, axis)
v5_2 = self.get_new_point(e1, v4, a2, axis)
if debug is True: # add debug verteces for the chosen verteces
dbgv1 = Vertex(v5_1)
particle.debug_verteces.add(dbgv1)
dbgv2 = Vertex(v5_2)
particle.debug_verteces.add(dbgv2)
dbgv3 = Vertex(E1.coord)
particle.debug_verteces.add(dbgv3)
# pick closest to centroid
if np.linalg.norm(v5_1 - particle.centroid) < np.linalg.norm(
v5_2 - particle.centroid):
new_point = v5_1
else:
new_point = v5_2
new_vertex = Vertex(new_point) # the final V5
return new_vertex
def add2(self, particle, template,
adding_edge): # returns a candidate new vertex coord
angle1 = template.angle1
angle2 = template.angle2
stem_length = template.stem_length
assert (self.full == False)
# V4,V5,V6,V7
#
# V1-------V3
# / \ /
# / \ /
# V7------V2
#
# E1 is midpoint of V1,V2 (the adding edge)
# V4 is the first point on the way to the final point (V7). V4 is 1 stem_length above E1, 90 degrees from the Plan(V1,V2,V3)
# V5 will be an intermeidate point after the first rotation
# A 90 degree rotation of V5 each way will give V6, a test point to see which direction to rotate
# A final rotation of V5 will yield V7
E1 = adding_edge
(V1, V2) = adding_edge.verteces
V3 = self.opposite_vertex(adding_edge)
# debug
V1.atom_name = "Be"
V2.atom_name = "Ba"
V3.atom_name = "Ca"
adding_edge.atom_name = "Na"
vec_V2V3 = V2.coord - V3.coord
vec_V2V1 = V2.coord - V1.coord
vec_normal = np.cross(vec_V2V1, vec_V2V3)
V4_1 = E1.coord + (stem_length * unit_vector(vec_normal))
V4_2 = E1.coord - (stem_length * unit_vector(vec_normal))
dbgv = Vertex(V4_1)
dbgv.atom_name = "S"
particle.debug_verteces.add(dbgv)
dbgv = Vertex(V4_2)
dbgv.atom_name = "O"
particle.debug_verteces.add(dbgv)
dist1 = np.linalg.norm(V4_1 - particle.centroid)
dist2 = np.linalg.norm(V4_2 - particle.centroid)
if (dist1 >= dist2): # take the value further from centroid
V4 = V4_1
print("chose V4_1")
else:
V4 = V4_2
print("chose V4_2")
# set origin to E1
e1 = E1.coord
v1 = V1.coord - e1
v2 = V2.coord - e1
v3 = V3.coord - e1
v4 = V4 - e1
#########
def get_new_point(
reference_point, start_point, angle, axis
): # where angle is radians of the stem length (edge to opposite vertex)
M = rotation_matrix(angle, axis)
new_point = np.dot(start_point, M[:3, :3].T)
return new_point + reference_point
# get two new points with consistent angle for first rotation
axis = E1.coord - V3.coord
axis = unit_vector(axis)
V5_1 = get_new_point(e1, v4, angle1, axis)
V5_2 = get_new_point(e1, v4, (2 * np.pi) - angle1, axis)
dbgv = Vertex(V5_1)
dbgv.atom_name = "Mg"
particle.debug_verteces.add(dbgv)
dbgv = Vertex(V5_2)
dbgv.atom_name = "Al"
particle.debug_verteces.add(dbgv)
#perform a test rotation of V5 to get V6 and determine where to rotate towards for V7
axis = V2.coord - V1.coord
V5 = V5_1 # chooseing V5_1
v5 = V5 - e1 # at some point will need to choose which V5 to use. (random choice?)
a1 = (0.5) * np.pi
a2 = (-0.5) * np.pi
V6_1 = get_new_point(e1, v5, a1, axis)
V6_2 = get_new_point(e1, v5, a2, axis)
dbgv = Vertex(V6_1)
dbgv.atom_name = "Cl"
particle.debug_verteces.add(dbgv)
dbgv = Vertex(V6_2)
dbgv.atom_name = "Na"
particle.debug_verteces.add(dbgv)
dist1 = np.linalg.norm(V6_1 - V3.coord)
dist2 = np.linalg.norm(V6_2 - V3.coord)
print(dist1, dist2)
if dist1 > dist2:
V6 = V6_1
print("chose V6_1")
sign = 1
else:
V6 = V6_2
print("chose V6_2")
sign = -1
# make final rotation of V5 to V7
a1 = (0.5 + (sign * angle2)) * np.pi
a2 = (1 + (sign * angle2)) * np.pi
V7_1 = get_new_point(e1, v5, a1, axis)
V7_2 = get_new_point(e1, v5, a2, axis)
dist1 = np.linalg.norm(V6 - V7_1)
dist2 = np.linalg.norm(V6 - V7_2)
if dist1 < dist2:
V7 = V7_1
else:
V7 = V7_2
dbgv = Vertex(V7_1)
dbgv.atom_name = "Mn"
particle.debug_verteces.add(dbgv)
dbgv = Vertex(V7_2)
dbgv.atom_name = "Li"
particle.debug_verteces.add(dbgv)
new_vertex = Vertex(V7)
return new_vertex # returns the adding_edge, new_vertex
def add(self, particle, template,
adding_edge): # returns a candidate new vertex coord
#print("Adding trimer template type:",template.template_type)
angle = template.angle
stem_length = template.stem_length
assert (self.full == False)
# edge = adding_edge
# opposite_vertex = self.opposite_vertex(edge)
# v1 = edge.vertex1.coord
# v2 = edge.vertex2.coord
# v3 = opposite_vertex.coord
# v4 = self.coord
# a = edge.coord
#
#print("adding template type:",template.template_type)
E1 = adding_edge
(V1, V2) = adding_edge.verteces
V3 = self.opposite_vertex(adding_edge)
# # maybe extend length a bit along the V1V3 vector to get the appropriate stem length for the template
vec = E1.coord - V3.coord
V1V3_dist = np.linalg.norm(V3.coord - E1.coord)
#print(V1V3_dist)
d = stem_length / V1V3_dist # this does not extend
#print(V1V3_dist,stem_length,d)
extended_point1 = E1.coord - d * vec
extended_point2 = E1.coord + d * vec
dist1 = np.linalg.norm(extended_point1 - V3.coord)
dist2 = np.linalg.norm(extended_point2 - V3.coord)
if dist1 < dist2:
extended_point = extended_point1
else:
extended_point = extended_point2
# set origin to E1
e1 = E1.coord
v1 = V1.coord - e1
v2 = V1.coord - e1
v3 = V3.coord - e1
extended_point = extended_point - e1
#########
def get_new_point(
reference_point, start_point, angle, axis
): # where angle is radians of the stem length (edge to opposite vertex)
M = rotation_matrix(angle, axis)
new_point = np.dot(start_point, M[:3, :3].T)
return new_point + reference_point
# get two new points with consistent angle for first rotation
axis = V1.coord - V2.coord
axis = unit_vector(axis)
p1 = get_new_point(e1, extended_point, angle, axis)
p2 = get_new_point(e1, extended_point, -1 * angle, axis)
#debug
# dbgv1 = Vertex(p1)
# dbgv2 = Vertex(p2)
# particle.debug_verteces.add(dbgv1)
# particle.debug_verteces.add(dbgv2)
# pick closest to centroid
if np.linalg.norm(p1 - particle.centroid) < np.linalg.norm(
p2 - particle.centroid):
new_point = p1
else:
new_point = p2
new_vertex = Vertex(new_point)
return new_vertex # returns the adding_edge, new_vertex
def check_clash_tops(self, adding_edge, candidate_vertex):
adding_trimer = next(iter(adding_edge.trimers))
# make tops
(v1, v2) = adding_edge.verteces
v3 = candidate_vertex
n1 = v2.coord - v1.coord
n2 = v3.coord - v1.coord
vec = np.cross(n1, n2)
d = 25
center = (adding_edge.coord + candidate_vertex.coord) / 2
top1 = Top(center + d * unit_vector(vec))
top2 = Top(center - d * unit_vector(vec))
new_tops = {top1, top2}
# get a set of trimers to check for clashes with (not closesly connected to current trimer)
check_trimers = self.exclude_nearby_trimers(adding_trimer)
# check for clashes
top_clash_flag = False
rejection = None
for trimer in check_trimers:
if top_clash_flag is False:
(v1, v2, v3) = trimer.verteces
(t1, t2) = trimer.tops
results = [
] # a list of True false, if any are True, there was a clash
toltt = self.clash_tolerance_top_top
toltv = self.clash_tolerance_top_vertex
if results.append(
self.check_clash_between_two_objects(top1, v1, toltv)):
top_clash_flag = True
elif results.append(
self.check_clash_between_two_objects(top1, v2, toltv)):
top_clash_flag = True
elif results.append(
self.check_clash_between_two_objects(top1, v3, toltv)):
top_clash_flag = True
elif results.append(
self.check_clash_between_two_objects(top1, t1, toltt)):
top_clash_flag = True
elif results.append(
self.check_clash_between_two_objects(top1, t2, toltt)):
top_clash_flag = True
elif results.append(
self.check_clash_between_two_objects(top2, v1, toltv)):
top_clash_flag = True
elif results.append(
self.check_clash_between_two_objects(top2, v2, toltv)):
top_clash_flag = True
elif results.append(
self.check_clash_between_two_objects(top2, v3, toltv)):
top_clash_flag = True
elif results.append(
self.check_clash_between_two_objects(top2, t1, toltt)):
top_clash_flag = True
elif results.append(
self.check_clash_between_two_objects(top2, t2, toltt)):
top_clash_flag = True
if top_clash_flag is True:
rejection = Rejection(self.particle.timestep, 3)
return top_clash_flag, rejection
def compute_error(trial, descriptor_type, niter, percentile=99):
src_scene_root = "/Users/isa/Experiments/reg3d_eval/downtown_dan/trial_" +str(trial);
src_features_dir = "/Users/isa/Experiments/reg3d_eval/downtown_dan/trial_" +str(trial)+ "/" + descriptor_type + "_30"
#read "geo-transformatiom"
#************GEO**************"
Tfile = "/data/lidar_providence/downtown_offset-1-financial-Hs.txt"
Tfis = open(Tfile, 'r')
lines=[];
lines = Tfis.readlines();
scale_geo = float(lines[0])
Ss_geo = tf.scale_matrix(scale_geo);
quat_line = lines[1].split(" ");
quat_geo = np.array([float(quat_line[3]), float(quat_line[0]), float(quat_line[1]), float(quat_line[2])])
Rs_geo = tf.quaternion_matrix(quat_geo);
trans_line = lines[2].split(" ");
trans_geo = np.array([float(trans_line[0]), float(trans_line[1]), float(trans_line[2])]);
Tfis.close();
Hs_geo = Rs_geo.copy();
Hs_geo[:3, 3] = trans_geo[:3]
Hs_geo = Ss_geo.dot(Hs_geo)
LOG.debug( "\n************Geo************** \n Scale: \n%s \nR:\n%s \nT:\n%s \nH:\n%s", Ss_geo, Rs_geo, trans_geo, Hs_geo)
#************Hs**************#
#read source to target "Ground Truth" Transformation
Tfile = src_scene_root + "/Hs.txt";
Tfis = open(Tfile, 'r')
lines=[];
lines = Tfis.readlines();
scale = float(lines[0])
Ss = tf.scale_matrix(scale);
quat_line = lines[1].split(" ");
quat = tf.unit_vector(np.array([float(quat_line[3]), float(quat_line[0]), float(quat_line[1]), float(quat_line[2])]))
Hs = tf.quaternion_matrix(quat);
trans_line = lines[2].split(" ");
Ts = np.array([float(trans_line[0]), float(trans_line[1]), float(trans_line[2])]);
Tfis.close();
Rs = Hs.copy()[:3,:3];
Hs[:3, 3] = Ts[:3]
Hs=Ss.dot(Hs)
Rs = Rs;
Ts = Ts;
LOG.debug( "\n************Hs************** \n R:\n%s \nT:\n%s \nH:\n%s", Rs, Ts, Hs)
#************Hs IA**************
#read source to target "Initial Alignment" Transformation
Tfile = src_features_dir + "/ia_transformation_" + str(percentile) + "_" + str(niter) + ".txt";
Tfis = open(Tfile, 'r')
Hs_ia = np.genfromtxt(Tfis, skip_header=1, usecols={0,1,2,3} );
Tfis.close()
Tfis = open(Tfile, 'r')
Ss_ia=np.genfromtxt(Tfis, skip_footer=4, usecols={0} );
Tfis.close()
Rs_ia = Hs_ia[:3,:3]*(1.0/Ss_ia)
Ts_ia = Hs_ia[:3,3]*(1.0/Ss_ia)
LOG.debug( "\n************Hs IA************** \n R:\n%s \nT:\n%s \nH:\n%s", Rs_ia, Ts_ia, Hs_ia)
#Initial Aligment errors
#Rotation error - half angle between the normalized quaterions
quat_ia = tf.unit_vector(tf.quaternion_from_matrix(Rs_ia));
Rs_error_ia_norm = math.acos(abs(np.dot(quat_ia, quat)));
#Translation error
# x = Rs_ia*x_ia + Ts_ia = Rs_ia(x_ia + np.dot(Rs_ia.T(), Ts_ia)
# np.dot(Rs_ia.T(), Ts_ia) correspond to trans on ia coordinate system
Ts_error_ia = (Rs_ia.T).dot(Ts_ia) - (Rs.T).dot(Ts)
Ts_error_ia_norm = scale_geo*scale*LA.norm(Ts_error_ia)
LOG.debug( "Error (R,T) %s , %s ", Rs_error_ia_norm , Ts_error_ia_norm )
#read source to target "Initial Alignment" Transformation
#************Hs ICP**************
Tfile = src_features_dir + "/icp_transformation_" + str(percentile) + "_" + str(niter) + ".txt";
Tfis = open(Tfile, 'r')
Hs_icp = np.genfromtxt(Tfis, usecols={0,1,2,3});
Tfis.close()
Hs_icp = Hs_icp.dot(Hs_ia)
Rs_icp = Hs_icp[:3,:3]*(1.0/Ss_ia)
Ts_icp = Hs_icp[:3,3]*(1.0/Ss_ia)
LOG.debug( "\n************Hs ICP************** \n R:\n%s \nT:\n%s \nH:\n%s", Rs_icp, Ts_icp, Hs_icp)
#ICP errors
#Rotation error - half angle between the normalized quaterions
quat_icp = tf.unit_vector(tf.quaternion_from_matrix(Rs_icp));
Rs_error_icp_norm = math.acos(abs(np.dot(quat_icp, quat)));
#Translation error
# x = Rs_ia*x_ia + Ts_ia = Rs_ia(x_ia + np.dot(Rs_ia.T(), Ts_ia)
# np.dot(Rs_ia.T(), Ts_ia) correspond to trans on ia coordinate system
Ts_error_icp = (Rs_icp.T).dot(Ts_icp) - (Rs.T).dot(Ts)
Ts_error_icp_norm = scale_geo*scale*LA.norm(Ts_error_icp)
LOG.debug( "Error (R,T) %s , %s ", Rs_error_icp_norm , Ts_error_icp_norm )
#read source to target "Initial Alignment" Transformation
#************Hs ICP Nprmals**************
Tfile = src_features_dir + "/icp_transformation_" + str(percentile) + "_" + str(niter) + "_n.txt";
Tfis = open(Tfile, 'r')
Hs_icn = np.genfromtxt(Tfis, usecols={0,1,2,3});
Tfis.close()
Hs_icn = Hs_icn.dot(Hs_ia)
Rs_icn = Hs_icn[:3,:3]*(1.0/Ss_ia)
Ts_icn = Hs_icn[:3,3]*(1.0/Ss_ia)
LOG.debug( "\n************Hs ICP Normals************** \n R:\n%s \nT:\n%s \nH:\n%s", Rs_icn, Ts_icn, Hs_icn)
#ICP errors
#Rotation error - half angle between the normalized quaterions
quat_icn = tf.unit_vector(tf.quaternion_from_matrix(Rs_icn));
Rs_error_icn_norm = math.acos(abs(np.dot(quat_icn, quat)));
#Translation error
# x = Rs_ia*x_ia + Ts_ia = Rs_ia(x_ia + np.dot(Rs_ia.T(), Ts_ia)
# np.dot(Rs_ia.T(), Ts_ia) correspond to trans on ia coordinate system
Ts_error_icn = (Rs_icn.T).dot(Ts_icn) - (Rs.T).dot(Ts)
Ts_error_icn_norm = scale_geo*scale*LA.norm(Ts_error_icn)
LOG.debug( "Error (R,T) %s , %s ", Rs_error_icn_norm , Ts_error_icn_norm )
ICP_error = np.array([Rs_error_icp_norm, Ts_error_icp_norm])
ICN_error = np.array([Rs_error_icn_norm, Ts_error_icn_norm]);
# import code; code.interact(local=locals())
return ICP_error,ICN_error
def __init__(self, T):
if ( type(T) == str):
print "Loading Transformation from file: " + T
Tfile = T
#Load transformation
Tfis = open(Tfile, 'r')
lines = []
lines = Tfis.readlines()
format = len(lines)
Tfis.seek(0) #reset file pointer
if not (format==3 or format==4 or format==5) :
raise ValueError("Wrong number of lines in file")
# import code; code.interact(local=locals())
if format == 3:
"""Handles processing a ground truth transfomation
File saved using vxl format
x' = s *(Rx + T)
scale
quaternion
translation """
print("Reading format 3")
self.scale = float(lines[0])
self.Ss = tf.scale_matrix(self.scale)
quat_line = lines[1].split(" ")
self.quat = tf.unit_vector(np.array([float(quat_line[3]),
float(quat_line[0]),
float(quat_line[1]),
float(quat_line[2])]))
self.Hs = tf.quaternion_matrix(self.quat)
trans_line = lines[2].split(" ")
self.Ts = np.array([float(trans_line[0]),
float(trans_line[1]),
float(trans_line[2])])
Tfis.close()
self.Rs = self.Hs.copy()[:3, :3]
self.Hs[:3, 3] = self.Ts[:3]
self.Hs = self.Ss.dot(self.Hs) # to add again
if format == 4 :
"""If the transformation was saved as:
H (4x4) - = [S*R|S*T]
"""
print("Reading format 4")
self.Hs = np.genfromtxt(Tfile, usecols={0, 1, 2, 3})
Tfis.close()
scale, shear, angles, trans, persp = tf.decompose_matrix(self.Hs)
self.scale = scale[0] # assuming isotropic scaling
self.Rs = self.Hs[:3, :3] * (1.0 / self.scale)
self.Ts = self.Hs[:3, 3] * (1.0 / self.scale)
self.quat = tf.quaternion_from_euler(angles[0], angles[1], angles[2])
if format==5:
"""If the transformation was saved as:
scale
H (4x4) - = [S*R|S*T]
"""
print("Reading format 5")
self.Hs = np.genfromtxt(Tfis, skip_header=1, usecols={0, 1, 2, 3})
Tfis.close()
Tfis = open(Tfile, 'r')
self.scale = np.genfromtxt(Tfis, skip_footer=4, usecols={0})
Tfis.close()
self.Rs = self.Hs[:3, :3] * (1.0 / self.scale)
self.Ts = self.Hs[:3, 3] * (1.0 / self.scale)
scale, shear, angles, trans, persp = tf.decompose_matrix(self.Hs)
self.quat = tf.quaternion_from_euler(angles[0], angles[1], angles[2])
print "Debugging translation:"
print self.Ts
print trans/self.scale
elif (type(T) == np.ndarray):
print "Loading Transformation array"
self.Hs = T
scale, shear, angles, trans, persp = tf.decompose_matrix(T)
self.scale =scale[0]
self.Rs = self.Hs[:3, :3] * (1.0 / self.scale)
self.Ts = self.Hs[:3, 3] * (1.0 / self.scale)
self.quat = tf.quaternion_from_euler(angles[0], angles[1], angles[2])
print "Debugging translation:"
print self.Ts
print trans/self.scale
# self.Rs = tf.quaternion_matrix(self.quat)
# self.Ts = trans / self.scale
print self.Hs
result_validity.append(False)
for i in range(len(lst)):
# Identify class id and compose dictionary key
class_id = int(lst[i]) - 1
key = class_names[class_id] + '_' + video_id
result_string = invalid_result_placeholder
if result_validity[i] == True:
# Convert quaternion to axis angle
q = my_result[i][0:4]
p = my_result[i][4:7]
np_q = np.quaternion(q[0], q[1], q[2], q[3])
axis_angle = quaternion.as_rotation_vector(np_q)
axis_normalized = unit_vector(axis_angle)
# Compose output vector as "<x y z> <axis> <angle> <6 null velocity vector> <frame id>"
result = [
p[0], p[1], p[2], axis_normalized[0], axis_normalized[1],
axis_normalized[2],
np.linalg.norm(axis_angle)
]
result_string = '{:s}'.format(' '.join(
['{:.7f}'.format(x)
for x in result[:7]])) + ' 0.0 0.0 0.0 0.0 0.0 0.0'
# Write on file
try:
results_files[key]
except KeyError:
def transformation_error(**kwargs):
root_dir = kwargs.get('root_dir')
descriptor_type = kwargs.get('descriptor_type')
radius = kwargs.get('radius', 30)
percentile = kwargs.get('percentile', 99)
nr_iterations = kwargs.get('nr_iterations', 500)
aux_output_string = kwargs.get('aux_output_string', "")
descriptor_string = kwargs.get('descriptor_string', "descriptors")
gt_fname = kwargs.get('gt_fname', "Hs.txt")
geo_tfile = kwargs.get('geo_tfile', "")
compute_scale = kwargs.get('compute_scale', False)
src_features_dir = root_dir + "/" + descriptor_type + "_" + str(radius)
#************GEO**************"
#load the geo tranformation
GEO = reg3d_T.geo_transformation(geo_tfile)
#************Hs**************#
#read source to target "Ground Truth" Transformation
Tfile = root_dir + "/" + gt_fname;
GT_Tform = reg3d_T.gt_transformation(Tfile)
#************PCL IA and ICP**************
#read source to target "Initial Alignment" Transformation
Tfile_ia = src_features_dir + "/ia_transformation_" + str(percentile) + "_" + str(nr_iterations) + "_" + aux_output_string + ".txt";
Tfile_icp = src_features_dir + "/icp_transformation_" + str(percentile) + "_" + str(nr_iterations) + "_" + aux_output_string + ".txt"
REG_Tform = reg3d_T.pcl_transformation(Tfile_ia, Tfile_icp, (not compute_scale))
#Initial Alignment errors
#Rotation error - half angle between the normalized quaternions
quat_ia = tf.unit_vector(tf.quaternion_from_matrix(REG_Tform.Rs_ia));
Rs_error_ia_norm = math.acos(abs(np.dot(quat_ia, GT_Tform.quat)))* 180/np.pi;
#Translation error
# x = REG_Tform.Rs_ia*x_ia + Ts_ia = REG_Tform.Rs_ia(x_ia + np.dot(REG_Tform.Rs_ia.T(), Ts_ia)
# np.dot(REG_Tform.Rs_ia.T(), Ts_ia) correspond to trans on ia coordinate system
Ts_error_ia = (REG_Tform.Rs_ia.T).dot(REG_Tform.Ts_ia) - (GT_Tform.Rs.T).dot(GT_Tform.Ts)
Ts_error_ia_norm = GEO.scale_geo*GT_Tform.scale*LA.norm(Ts_error_ia)
scale_error_ia = 1.0 - REG_Tform.scale_ia/GT_Tform.scale
print "Error (S,R,T)", scale_error_ia, Rs_error_ia_norm , Ts_error_ia_norm
#ICP errors
#Rotation error - half angle between the normalized quaternions
quat_icp = tf.unit_vector(tf.quaternion_from_matrix(REG_Tform.Rs_icp))
Rs_error_icp_norm = math.acos(abs(np.dot(quat_icp, GT_Tform.quat)))* 180/np.pi
#Translation error
# x = REG_Tform.Rs_ia*x_ia + Ts_ia = REG_Tform.Rs_ia(x_ia + np.dot(REG_Tform.Rs_ia.T(), Ts_ia)
# np.dot(REG_Tform.Rs_ia.T(), Ts_ia) correspond to trans on ia coordinate system
Ts_error_icp = (REG_Tform.Rs_icp.T).dot(REG_Tform.Ts_icp) - (GT_Tform.Rs.T).dot(GT_Tform.Ts)
Ts_error_icp_norm = GEO.scale_geo*GT_Tform.scale*LA.norm(Ts_error_icp)
scale_error_icp = 1.0 - (REG_Tform.scale_icp*REG_Tform.scale_ia)/GT_Tform.scale
print "Error (S, R,T)", scale_error_icp, Rs_error_icp_norm , Ts_error_icp_norm
IA_error = np.array([scale_error_ia, Rs_error_ia_norm, Ts_error_ia_norm])
ICP_error = np.array([scale_error_icp, Rs_error_icp_norm, Ts_error_icp_norm])
# import code; code.interact(local=locals())
return IA_error, ICP_error
def train(palette):
with tf.Graph().as_default():
with tf.device('/cpu:0'):
###################################################
### placeholders # triplet input
data_train_placeholder = tf.placeholder(tf.float32,
shape=(BATCH_SIZE,
NUM_POINT,
data_dim))
data_positive_train_placeholder = tf.placeholder(tf.float32,
shape=(BATCH_SIZE,
NUM_POINT,
data_dim))
data_negative_train_placeholder = tf.placeholder(tf.float32,
shape=(BATCH_SIZE,
NUM_POINT,
data_dim))
is_training_placeholder = tf.placeholder(tf.bool, shape=())
cluster_mean_placeholder = tf.placeholder(tf.float32,
shape=(cluster_num, 128))
indices_placeholder = tf.placeholder(tf.int64, shape=(BATCH_SIZE))
batch = tf.get_variable('batch', [],
initializer=tf.constant_initializer(0),
trainable=False)
bn_decay = get_bn_decay(batch)
tf.summary.scalar('bn_decay', bn_decay)
###################################################
### get training operator
learning_rate = get_learning_rate(batch)
tf.summary.scalar('learning_rate', learning_rate)
if OPTIMIZER == 'momentum':
optimizer = tf.train.MomentumOptimizer(learning_rate,
momentum=MOMENTUM)
elif OPTIMIZER == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate)
###################################################
tower_grads = []
feature_anchor_gpu = []
net_anchor_gpu = []
net_positive_gpu = []
net_negative_gpu = []
cluster_loss_gpu = []
contrastive_loss_gpu = []
total_loss_gpu = []
l2_loss_gpu = []
all_loss_gpu = []
### get model and loss
for i in range(NUM_GPUS):
with tf.variable_scope(tf.get_variable_scope(),
reuse=bool(i != 0)):
with tf.device('/gpu:%d' % (i)), tf.name_scope(
'gpu_%d' % (i)) as scope:
pc_patch = tf.slice(data_train_placeholder,
[i * DEVICE_BATCH_SIZE, 0, 0],
[DEVICE_BATCH_SIZE, -1, -1])
pc_patch_positive = tf.slice(
data_positive_train_placeholder,
[i * DEVICE_BATCH_SIZE, 0, 0],
[DEVICE_BATCH_SIZE, -1, -1])
pc_patch_negative = tf.slice(
data_negative_train_placeholder,
[i * DEVICE_BATCH_SIZE, 0, 0],
[DEVICE_BATCH_SIZE, -1, -1])
pc_features, pc_net, _ = MODEL.get_model(
pc_patch, is_training_placeholder, bn_decay)
tf.get_variable_scope().reuse_variables()
pc_features_positive, pc_net_positive, _ = MODEL.get_model(
pc_patch_positive, is_training_placeholder,
bn_decay)
pc_features_negative, pc_net_negative, _ = MODEL.get_model(
pc_patch_negative, is_training_placeholder,
bn_decay)
pc_indices = tf.slice(indices_placeholder,
[i * DEVICE_BATCH_SIZE],
[DEVICE_BATCH_SIZE])
pc_cluster_loss, _, _, pc_contrastive_loss, pc_total_loss = MODEL.get_loss(
pc_net, pc_net_positive, pc_net_negative,
cluster_mean_placeholder, pc_indices)
update_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
update_ops = tf.group(*update_op)
with tf.control_dependencies([update_ops]):
pc_l2_loss = tf.losses.get_regularization_loss()
pc_l2_loss = 0.00001 * pc_l2_loss
pc_all_loss = pc_total_loss + pc_l2_loss
grads = optimizer.compute_gradients(pc_all_loss)
tower_grads.append(grads)
cluster_loss_gpu.append(pc_cluster_loss)
contrastive_loss_gpu.append(pc_contrastive_loss)
total_loss_gpu.append(pc_total_loss)
l2_loss_gpu.append(pc_l2_loss)
all_loss_gpu.append(pc_all_loss)
feature_anchor_gpu.append(pc_features)
net_anchor_gpu.append(pc_net)
net_positive_gpu.append(pc_net_positive)
net_negative_gpu.append(pc_net_negative)
cluster_loss = tf.reduce_mean(cluster_loss_gpu)
contrastive_loss = tf.reduce_mean(contrastive_loss_gpu)
total_loss = tf.reduce_mean(total_loss_gpu)
l2_loss = tf.reduce_mean(l2_loss_gpu)
all_loss = tf.reduce_mean(all_loss_gpu)
tf.summary.scalar('cluster_loss', cluster_loss)
tf.summary.scalar('contrastive_loss', contrastive_loss)
tf.summary.scalar('l2_loss', l2_loss)
tf.summary.scalar('total loss', all_loss)
feature_anchor = tf.concat(feature_anchor_gpu, 0)
net_anchor = tf.concat(net_anchor_gpu, 0)
net_positive = tf.concat(net_positive_gpu, 0)
net_negative = tf.concat(net_negative_gpu, 0)
grads = average_gradients(tower_grads)
apply_gradient_op = optimizer.apply_gradients(grads,
global_step=batch)
variable_averages = tf.train.ExponentialMovingAverage(
bn_decay, batch)
variables_averages_op = variable_averages.apply(
tf.trainable_variables())
train_op = tf.group(apply_gradient_op, variables_averages_op)
saver = tf.train.Saver(max_to_keep=11)
# Create a session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
config.log_device_placement = False
sess = tf.Session(config=config)
# Add summary writers
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(os.path.join(LOG_DIR, 'train'),
sess.graph)
# Init variables
init = tf.global_variables_initializer()
sess.run(init, {is_training_placeholder: True})
restore_epoch, checkpoint_path = model_utils.pre_load_checkpoint(
LOG_DIR)
global LOG_FOUT
if restore_epoch == 0:
LOG_FOUT = open(os.path.join(LOG_DIR, 'log_train.txt'), 'w')
LOG_FOUT.write(str(socket.gethostname()) + '\n')
LOG_FOUT.write(str(FLAGS) + '\n')
else:
LOG_FOUT = open(os.path.join(LOG_DIR, 'log_train.txt'), 'a')
saver.restore(sess, checkpoint_path)
ops = {
'pointclouds_pl': data_train_placeholder,
'pointclouds_pl_positive': data_positive_train_placeholder,
'pointclouds_pl_negative': data_negative_train_placeholder,
'is_training_pl': is_training_placeholder,
'buffer': cluster_mean_placeholder,
'indics': indices_placeholder,
'feature_anchor': feature_anchor,
'net_anchor': net_anchor,
'net_positive': net_positive,
'net_negative': net_negative,
'cluster_loss': cluster_loss,
'contrastive_loss': contrastive_loss,
'total_loss': total_loss,
'l2_loss': l2_loss,
'all_loss': all_loss,
'train_op': train_op,
'merged': merged,
'step': batch,
}
feature_buffer = np.random.rand(file_size, 128)
feature_buffer = trafo.unit_vector(feature_buffer, axis=1) # normalize
for epoch in tqdm(range(restore_epoch, MAX_EPOCH)):
# for epoch in tqdm(range(MAX_EPOCH)):
print('******* EPOCH %03d *******' % (epoch))
sys.stdout.flush()
## clustering
# cluster the memory buffer
cluster_pred = SpectralClustering(
n_clusters=cluster_num, gamma=1).fit_predict(feature_buffer)
# cluster_pred = KMeans(n_clusters=cluster_num).fit_predict(feature_buffer)
# cluster_pred = AgglomerativeClustering(n_clusters=cluster_num).fit_predict(feature_buffer)
cluster_mean = np.zeros(shape=(cluster_num, 128))
for cluster_idx in range(cluster_num):
indices = np.where(cluster_pred == cluster_idx)[0]
cluster_avg = np.mean(feature_buffer[indices, :], axis=0)
cluster_mean[cluster_idx, :] = cluster_avg
cluster_mean = trafo.unit_vector(cluster_mean, axis=1)
train_one_epoch(sess, cluster_mean, cluster_pred, data, ops,
train_writer)
## get the feature buffer after each epoch
list = range(pts_num)
seed = random.sample(list, NUM_POINT)
data_2048 = data[:, seed, :]
num_batches = file_size // BATCH_SIZE
feature_buffer = np.zeros(shape=(file_size, 128))
for batch_idx in range(num_batches + 1):
if batch_idx != num_batches:
start_idx = batch_idx * BATCH_SIZE
end_idx = (batch_idx + 1) * BATCH_SIZE
else:
start_idx = file_size - BATCH_SIZE
end_idx = file_size
data_input = data_2048[start_idx:end_idx, :, :]
feed_dict = {
ops['pointclouds_pl']: data_input,
ops['is_training_pl']: False,
}
features = sess.run([ops['net_anchor']], feed_dict=feed_dict)
feature_buffer[start_idx:end_idx, :] = features[0]
if epoch % 50 == 0:
# start_time = time.time()
# tsne_visualization_color(feature_buffer, cluster_pred, palette, epoch)
# end_time = time.time()
# print 'tsne running time is %f' % (end_time-start_time)
# start_time = time.time()
# pca_visualization_color(feature_buffer, cluster_pred, palette, epoch)
# end_time = time.time()
# print 'pca running time is %f' % (end_time-start_time)
# if epoch % 10 == 0:
save_path = saver.save(sess,
os.path.join(LOG_DIR, "model.ckpt"),
global_step=epoch)
print("Model saved in file: %s" % save_path)
