def sample_landmark(class_, landmarks, trajector):
distances = array([lmk.distance_to(trajector.representation) for lmk in landmarks])
scores = 1.0/(array(distances)**1.5 + class_.epsilon)
scores[distances == 0] = 0
lm_probabilities = scores/sum(scores)
index = lm_probabilities.cumsum().searchsorted( random.sample(1) )[0]
return index
def sample_landmark(self, landmarks, trajector):
""" Weight by inverse of distance to landmark center and choose probabilistically """
epsilon = 0.02
distances = array(
[
trajector.distance_to(lmk.representation)
if not (
isinstance(lmk.representation, RectangleRepresentation)
and lmk.representation.contains(trajector.representation)
)
else 9 * epsilon
for lmk in landmarks
]
)
# distances = array([trajector.distance_to( lmk )
# if not (isinstance(lmk.representation,RectangleRepresentation) and lmk.representation.contains(trajector))
# else min(poly_to_vec_distance(lmk.representation.get_geometry().to_polygon(), trajector.representation.location),lmk.representation.middle.distance_to(trajector.representation.location))
# for lmk in landmarks])
# scores = 1.0/(distances + epsilon)**0.5
std = 0.1
scores = exp(-(distances / std) ** 2)
lm_probabilities = scores / sum(scores)
index = lm_probabilities.cumsum().searchsorted(random.sample(1))[0]
sampled_landmark = landmarks[index]
head_on = self.get_head_on_viewpoint(sampled_landmark)
self.set_orientations(sampled_landmark, head_on)
return sampled_landmark, lm_probabilities[index], self.get_entropy(lm_probabilities), head_on
def sample_landmark(class_, landmarks, trajector):
distances = array(
[lmk.distance_to(trajector.representation) for lmk in landmarks])
scores = 1.0 / (array(distances)**1.5 + class_.epsilon)
scores[distances == 0] = 0
lm_probabilities = scores / sum(scores)
index = lm_probabilities.cumsum().searchsorted(random.sample(1))[0]
return index
def sample_point_trajector(self, bounding_box, relation, perspective, landmark, step=0.02):
"""
Sample a point of interest given a relation and landmark.
"""
probs, points = self.get_probabilities_box(bounding_box, relation, perspective, landmark)
probs /= probs.sum()
index = probs.cumsum().searchsorted( random.sample(1) )[0]
return Landmark( 'point', Vec2( *points[index] ), None, Landmark.POINT )
def sample_relation(class_, perspective, sampled_landmark, trajector):
rel_scores = []
rel_instances = []
for relation in class_.relations:
rel_instances.append( relation(perspective, sampled_landmark, trajector) )
rel_scores.append( rel_instances[-1].is_applicable() )
rel_scores = array(rel_scores)
rel_probabilities = rel_scores/sum(rel_scores)
index = rel_probabilities.cumsum().searchsorted( random.sample(1) )[0]
return rel_instances[index]
def sample_relation(class_, perspective, sampled_landmark, trajector):
rel_scores = []
rel_instances = []
for relation in class_.relations:
rel_instances.append(
relation(perspective, sampled_landmark, trajector))
rel_scores.append(rel_instances[-1].is_applicable())
rel_scores = array(rel_scores)
rel_probabilities = rel_scores / sum(rel_scores)
index = rel_probabilities.cumsum().searchsorted(random.sample(1))[0]
return rel_instances[index]
def sample_point_trajector(self,
bounding_box,
relation,
perspective,
landmark,
step=0.02):
"""
Sample a point of interest given a relation and landmark.
"""
probs, points = self.get_probabilities_box(bounding_box, relation,
perspective, landmark)
probs /= probs.sum()
index = probs.cumsum().searchsorted(random.sample(1))[0]
return Landmark('point', Vec2(*points[index]), None, Landmark.POINT)
def sample_relation(self, trajector, bounding_box, perspective, landmark, step=0.02, usebest=False):
"""
Sample a relation given a trajector and landmark.
Evaluate each relation and probabilisticaly choose the one that is likely to
generate the trajector given a landmark.
"""
rel_probabilities, rel_classes = self.all_relation_probs(trajector, bounding_box, perspective, landmark, step)
if usebest:
index = index_max(rel_probabilities)
else:
index = categorical_sample(rel_probabilities)
index = rel_probabilities.cumsum().searchsorted( random.sample(1) )[0]
return rel_classes[index], rel_probabilities[index], self.get_entropy(rel_probabilities)
def evaluate_all(self):
epsilon = 1e-6
probs = []
for dist in self.distance_classes:
for degree in self.degree_classes:
p = Measurement.get_applicability(self.distance, dist, degree) + epsilon
probs.append([p, degree, dist])
ps, degrees, dists = zip(*probs)
ps = array(ps)
ps /= sum(ps)
index = ps.cumsum().searchsorted( random.sample(1) )[0]
return probs[index]
def evaluate_all(self):
epsilon = 1e-6
probs = []
for dist in self.distance_classes:
for degree in self.degree_classes:
p = Measurement.get_applicability(self.distance, dist,
degree) + epsilon
probs.append([p, degree, dist])
ps, degrees, dists = zip(*probs)
ps = array(ps)
ps /= sum(ps)
index = ps.cumsum().searchsorted(random.sample(1))[0]
return probs[index]
def sample_relation(self, trajector, bounding_box, perspective, landmark, step=0.02):
"""
Sample a relation given a trajector and landmark.
Evaluate each relation and probabilisticaly choose the one that is likely to
generate the trajector given a landmark.
"""
rel_scores = []
rel_classes = []
for s in [DistanceRelationSet, ContainmentRelationSet]:
for rel in s.relations:
rel_scores.append(
self.evaluate_trajector_likelihood(trajector, bounding_box, rel, perspective, landmark, step)
)
rel_classes.append(rel)
ori_rel_scores = []
for rel in OrientationRelationSet.relations:
p = self.evaluate_trajector_likelihood(trajector, bounding_box, rel, perspective, landmark, step)
if p > 0:
ori_rel_scores.append((p, rel))
if len(ori_rel_scores) > 1:
assert len(ori_rel_scores) == 2
dists = []
for p, rel in ori_rel_scores:
dists.append([rel(perspective, landmark, trajector).measurement.distance, p, rel])
dists = sorted(dists)
dists[0][1] *= dists[0][0] / dists[1][0]
rel_scores.append(dists[0][1])
rel_scores.append(dists[1][1])
rel_classes.append(dists[0][2])
rel_classes.append(dists[1][2])
rel_scores = array(rel_scores)
rel_probabilities = rel_scores / sum(rel_scores)
index = rel_probabilities.cumsum().searchsorted(random.sample(1))[0]
return rel_classes[index], rel_probabilities[index], self.get_entropy(rel_probabilities)
def sample_relation(self,
trajector,
bounding_box,
perspective,
landmark,
step=0.02,
usebest=False):
"""
Sample a relation given a trajector and landmark.
Evaluate each relation and probabilisticaly choose the one that is likely to
generate the trajector given a landmark.
"""
rel_probabilities, rel_classes = self.all_relation_probs(
trajector, bounding_box, perspective, landmark, step)
if usebest:
index = index_max(rel_probabilities)
else:
index = categorical_sample(rel_probabilities)
index = rel_probabilities.cumsum().searchsorted(random.sample(1))[0]
return rel_classes[index], rel_probabilities[index], self.get_entropy(
rel_probabilities)