class EAOScore(Analysis):
eaocurve = Include(EAOCurve)
low = Integer()
high = Integer()
@property
def title(self):
return "EAO analysis"
def describe(self):
return Measure("Expected average overlap", "EAO", 0, 1,
Sorting.DESCENDING),
def compatible(self, experiment: Experiment):
return isinstance(experiment, SupervisedExperiment)
def dependencies(self):
return self.eaocurve,
def compute(self, experiment: Experiment, trackers: List[Tracker],
sequences: List[Sequence], dependencies: List[Grid]) -> Grid:
return dependencies[0].foreach(
lambda x, i, j: (float(np.mean(x[0][self.low:self.high + 1])), ))
@property
def axes(self):
return Axes.TRACKERS
class TrackerSorter(Attributee):
experiment = String(default=None)
analysis = String(default=None)
result = Integer(val_min=0, default=0)
def __call__(self, experiments, trackers, sequences):
if self.experiment is None or self.analysis is None:
return range(len(trackers))
experiment = next(
filter(lambda x: x.identifier == self.experiment, experiments),
None)
if experiment is None:
raise RuntimeError("Experiment not found")
analysis = next(
filter(lambda x: x.name == self.analysis, experiment.analyses),
None)
if analysis is None:
raise RuntimeError("Analysis not found")
future = analysis.commit(experiment, trackers, sequences)
result = future.result()
scores = [x[self.result] for x in result]
indices = [
i[0] for i in sorted(
enumerate(scores), reverse=True, key=lambda x: x[1])
]
return indices
class _Thresholds(SequenceAggregator):
resolution = Integer(default=100)
@property
def title(self):
return "Thresholds for tracking precision/recall"
def describe(self):
return None,
def compatible(self, experiment: Experiment):
return isinstance(experiment, UnsupervisedExperiment)
def dependencies(self):
return _ConfidenceScores(),
def aggregate(self, tracker: Tracker, sequences: List[Sequence],
results: Grid) -> Tuple[Any]:
thresholds = determine_thresholds(
itertools.chain(*[result[0] for result in results]),
self.resolution),
return thresholds,
class SequenceAccuracy(SeparableAnalysis):
burnin = Integer(default=10, val_min=0)
ignore_unknown = Boolean(default=True)
bounded = Boolean(default=True)
def compatible(self, experiment: Experiment):
return isinstance(experiment, MultiRunExperiment)
@property
def title(self):
return "Sequence accurarcy"
def describe(self):
return Measure("Accuracy", "AUC", 0, 1, Sorting.DESCENDING),
def subcompute(self, experiment: Experiment, tracker: Tracker,
sequence: Sequence, dependencies: List[Grid]) -> Tuple[Any]:
assert isinstance(experiment, MultiRunExperiment)
trajectories = experiment.gather(tracker, sequence)
if len(trajectories) == 0:
raise MissingResultsException()
cummulative = 0
for trajectory in trajectories:
accuracy, _ = compute_accuracy(trajectory.regions(), sequence,
self.burnin, self.ignore_unknown,
self.bounded)
cummulative = cummulative + accuracy
return cummulative / len(trajectories),
class Redetection(Transformer):
length = Integer(default=100, val_min=1)
initialization = Integer(default=5, val_min=1)
padding = Float(default=2, val_min=0)
scaling = Float(default=1, val_min=0.1, val_max=10)
def __call__(self, sequence: Sequence) -> Sequence:
chache_dir = self._cache.directory(self, arg_hash(sequence.name, **self.dump()))
if not os.path.isfile(os.path.join(chache_dir, "sequence")):
generated = InMemorySequence(sequence.name, sequence.channels())
size = (int(sequence.size[0] * self.scaling), int(sequence.size[1] * self.scaling))
initial_images = dict()
redetect_images = dict()
for channel in sequence.channels():
rect = sequence.frame(0).groundtruth().convert(RegionType.RECTANGLE)
halfsize = int(max(rect.width, rect.height) * self.scaling / 2)
x, y = rect.center()
image = Image.fromarray(sequence.frame(0).image())
box = (x - halfsize, y - halfsize, x + halfsize, y + halfsize)
template = image.crop(box)
initial = Image.new(image.mode, size)
initial.paste(image, (0, 0))
redetect = Image.new(image.mode, size)
redetect.paste(template, (size[0] - template.width, size[1] - template.height))
initial_images[channel] = initial
redetect_images[channel] = redetect
generated.append(initial_images, sequence.frame(0).groundtruth())
generated.append(redetect_images, sequence.frame(0).groundtruth().move(size[0] - template.width, size[1] - template.height))
write_sequence(chache_dir, generated)
source = VOTSequence(chache_dir, name=sequence.name)
mapping = [0] * self.initialization + [1] * (self.length - self.initialization)
return FrameMapSequence(source, mapping)
class MultiRunExperiment(Experiment):
repetitions = Integer(val_min=1, default=1)
early_stop = Boolean(default=True)
def _can_stop(self, tracker: Tracker, sequence: Sequence):
if not self.early_stop:
return False
trajectories = self.gather(tracker, sequence)
if len(trajectories) < 3:
return False
for trajectory in trajectories[1:]:
if not trajectory.equals(trajectories[0]):
return False
return True
def scan(self, tracker: Tracker, sequence: Sequence):
results = self.results(tracker, sequence)
files = []
complete = True
for i in range(1, self.repetitions + 1):
name = "%s_%03d" % (sequence.name, i)
if Trajectory.exists(results, name):
files.extend(Trajectory.gather(results, name))
elif self._can_stop(tracker, sequence):
break
else:
complete = False
break
return complete, files, results
def gather(self, tracker: Tracker, sequence: Sequence):
trajectories = list()
results = self.results(tracker, sequence)
for i in range(1, self.repetitions + 1):
name = "%s_%03d" % (sequence.name, i)
if Trajectory.exists(results, name):
trajectories.append(Trajectory.read(results, name))
return trajectories
class AccuracyRobustness(SeparableAnalysis):
sensitivity = Float(default=30, val_min=1)
burnin = Integer(default=10, val_min=0)
ignore_unknown = Boolean(default=True)
bounded = Boolean(default=True)
@property
def title(self):
return "AR analysis"
def describe(self):
return Measure("Accuracy", "A", minimal=0, maximal=1, direction=Sorting.DESCENDING), \
Measure("Robustness", "R", minimal=0, direction=Sorting.ASCENDING), \
Point("AR plot", dimensions=2, abbreviation="AR", minimal=(0, 0), \
maximal=(1, 1), labels=("Robustness", "Accuracy"), trait="ar"), \
None
def compatible(self, experiment: Experiment):
return isinstance(experiment, SupervisedExperiment)
def subcompute(self, experiment: Experiment, tracker: Tracker,
sequence: Sequence, dependencies: List[Grid]) -> Tuple[Any]:
trajectories = experiment.gather(tracker, sequence)
if len(trajectories) == 0:
raise MissingResultsException()
accuracy = 0
failures = 0
for trajectory in trajectories:
failures += count_failures(trajectory.regions())[0]
accuracy += compute_accuracy(trajectory.regions(), sequence,
self.burnin, self.ignore_unknown,
self.bounded)[0]
ar = (math.exp(-(float(failures) / len(trajectories)) *
self.sensitivity), accuracy / len(trajectories))
return accuracy / len(trajectories), failures / len(
trajectories), ar, len(trajectories[0])
class SupervisedExperiment(MultiRunExperiment):
skip_initialize = Integer(val_min=1, default=1)
skip_tags = List(String(), default=[])
failure_overlap = Float(val_min=0, val_max=1, default=0)
def execute(self,
tracker: Tracker,
sequence: Sequence,
force: bool = False,
callback: Callable = None):
results = self.results(tracker, sequence)
with self._get_runtime(tracker, sequence) as runtime:
for i in range(1, self.repetitions + 1):
name = "%s_%03d" % (sequence.name, i)
if Trajectory.exists(results, name) and not force:
continue
if self._can_stop(tracker, sequence):
return
trajectory = Trajectory(sequence.length)
frame = 0
while frame < sequence.length:
_, properties, elapsed = runtime.initialize(
sequence.frame(frame),
self._get_initialization(sequence, frame))
properties["time"] = elapsed
trajectory.set(frame, Special(Special.INITIALIZATION),
properties)
frame = frame + 1
while frame < sequence.length:
region, properties, elapsed = runtime.update(
sequence.frame(frame))
properties["time"] = elapsed
if calculate_overlap(
region, sequence.groundtruth(frame),
sequence.size) <= self.failure_overlap:
trajectory.set(frame, Special(Special.FAILURE),
properties)
frame = frame + self.skip_initialize
if self.skip_tags:
while frame < sequence.length:
if not [
t for t in sequence.tags(frame)
if t in self.skip_tags
]:
break
frame = frame + 1
break
else:
trajectory.set(frame, region, properties)
frame = frame + 1
if callback:
callback(i / self.repetitions)
trajectory.write(results, name)
class EAOCurve(TrackerSeparableAnalysis):
burnin = Integer(default=10, val_min=0)
bounded = Boolean(default=True)
@property
def title(self):
return "EAO Curve"
def describe(self):
return Plot("Expected Average Overlap",
"EAO",
minimal=0,
maximal=1,
trait="eao"),
def compatible(self, experiment: Experiment):
return isinstance(experiment, SupervisedExperiment)
def subcompute(self, experiment: Experiment, tracker: Tracker,
sequences: List[Sequence],
dependencies: List[Grid]) -> Tuple[Any]:
overlaps_all = []
weights_all = []
success_all = []
for sequence in sequences:
trajectories = experiment.gather(tracker, sequence)
if len(trajectories) == 0:
raise MissingResultsException()
for trajectory in trajectories:
overlaps = calculate_overlaps(
trajectory.regions(), sequence.groundtruth(),
(sequence.size) if self.bounded else None)
fail_idxs, init_idxs = locate_failures_inits(
trajectory.regions())
if len(fail_idxs) > 0:
for i in range(len(fail_idxs)):
overlaps_all.append(
overlaps[init_idxs[i]:fail_idxs[i]])
success_all.append(False)
weights_all.append(1)
# handle last initialization
if len(init_idxs) > len(fail_idxs):
# tracker was initilized, but it has not failed until the end of the sequence
overlaps_all.append(overlaps[init_idxs[-1]:])
success_all.append(True)
weights_all.append(1)
else:
overlaps_all.append(overlaps)
success_all.append(True)
weights_all.append(1)
return compute_eao_curve(overlaps_all, weights_all, success_all),
class InjectConfig(Attributee):
# Not implemented yet
placeholder = Integer(default=1)
class RealtimeConfig(Attributee):
grace = Integer(val_min=0, default=0)
fps = Float(val_min=0, default=20)
class AccuracyRobustness(SeparableAnalysis):
burnin = Integer(default=10, val_min=0)
grace = Integer(default=10, val_min=0)
bounded = Boolean(default=True)
threshold = Float(default=0.1, val_min=0, val_max=1)
@property
def title(self):
return "AR Analysis"
def describe(self):
return Measure("Accuracy", "A", minimal=0, maximal=1, direction=Sorting.DESCENDING), \
Measure("Robustness", "R", minimal=0, direction=Sorting.DESCENDING), \
Point("AR plot", dimensions=2, abbreviation="AR",
minimal=(0, 0), maximal=(1, 1), labels=("Robustness", "Accuracy"), trait="ar"), \
None, None
def compatible(self, experiment: Experiment):
return isinstance(experiment, MultiStartExperiment)
def subcompute(self, experiment: Experiment, tracker: Tracker,
sequence: Sequence, dependencies: List[Grid]) -> Tuple[Any]:
results = experiment.results(tracker, sequence)
forward, backward = find_anchors(sequence, experiment.anchor)
if not forward and not backward:
raise RuntimeError("Sequence does not contain any anchors")
robustness = 0
accuracy = 0
total = 0
for i, reverse in [(f, False) for f in forward] + [(f, True)
for f in backward]:
name = "%s_%08d" % (sequence.name, i)
if not Trajectory.exists(results, name):
raise MissingResultsException()
if reverse:
proxy = FrameMapSequence(sequence,
list(reversed(range(0, i + 1))))
else:
proxy = FrameMapSequence(sequence,
list(range(i, sequence.length)))
trajectory = Trajectory.read(results, name)
overlaps = calculate_overlaps(
trajectory.regions(), proxy.groundtruth(),
(proxy.size) if self.burnin else None)
grace = self.grace
progress = len(proxy)
for j, overlap in enumerate(overlaps):
if overlap <= self.threshold and not proxy.groundtruth(
j).is_empty():
grace = grace - 1
if grace == 0:
progress = j + 1 - self.grace # subtract since we need actual point of the failure
break
else:
grace = self.grace
robustness += progress # simplified original equation: len(proxy) * (progress / len(proxy))
accuracy += sum(overlaps[0:progress])
total += len(proxy)
ar = (robustness / total,
accuracy / robustness if robustness > 0 else 0)
return accuracy / robustness if robustness > 0 else 0, robustness / total, ar, robustness, len(
sequence)
class EAOCurves(SeparableAnalysis):
burnin = Integer(default=10, val_min=0)
grace = Integer(default=10, val_min=0)
bounded = Boolean(default=True)
threshold = Float(default=0.1, val_min=0, val_max=1)
high = Integer()
@property
def title(self):
return "EAO Curve"
def describe(self):
return Plot("Expected average overlap",
"EAO",
minimal=0,
maximal=1,
wrt="frames",
trait="eao"),
def compatible(self, experiment: Experiment):
return isinstance(experiment, MultiStartExperiment)
def subcompute(self, experiment: Experiment, tracker: Tracker,
sequence: Sequence, dependencies: List[Grid]) -> Tuple[Any]:
results = experiment.results(tracker, sequence)
forward, backward = find_anchors(sequence, experiment.anchor)
if len(forward) == 0 and len(backward) == 0:
raise RuntimeError("Sequence does not contain any anchors")
overlaps_all = []
success_all = []
for i, reverse in [(f, False) for f in forward] + [(f, True)
for f in backward]:
name = "%s_%08d" % (sequence.name, i)
if not Trajectory.exists(results, name):
raise MissingResultsException()
if reverse:
proxy = FrameMapSequence(sequence,
list(reversed(range(0, i + 1))))
else:
proxy = FrameMapSequence(sequence,
list(range(i, sequence.length)))
trajectory = Trajectory.read(results, name)
overlaps = calculate_overlaps(trajectory.regions(),
proxy.groundtruth(),
proxy.size if self.burnin else None)
grace = self.grace
progress = len(proxy)
for j, overlap in enumerate(overlaps):
if overlap <= self.threshold and not proxy.groundtruth(
j).is_empty():
grace = grace - 1
if grace == 0:
progress = j + 1 - self.grace # subtract since we need actual point of the failure
break
else:
grace = self.grace
success = True
if progress < len(overlaps):
# tracker has failed during this run
overlaps[progress:] = (len(overlaps) - progress) * [float(0)]
success = False
overlaps_all.append(overlaps)
success_all.append(success)
return compute_eao_partial(overlaps_all, success_all, self.high), 1
class MultiStartFragments(SeparableAnalysis):
burnin = Integer(default=10, val_min=0)
grace = Integer(default=10, val_min=0)
bounded = Boolean(default=True)
threshold = Float(default=0.1, val_min=0, val_max=1)
@property
def title(self):
return "Fragment Analysis"
def describe(self):
return Curve("Success",
2,
"Sc",
minimal=(0, 0),
maximal=(1, 1),
trait="points"), Curve("Accuracy",
2,
"Ac",
minimal=(0, 0),
maximal=(1, 1),
trait="points")
def compatible(self, experiment: Experiment):
return isinstance(experiment, MultiStartExperiment)
def subcompute(self, experiment: Experiment, tracker: Tracker,
sequence: Sequence, dependencies: List[Grid]) -> Tuple[Any]:
results = experiment.results(tracker, sequence)
forward, backward = find_anchors(sequence, experiment.anchor)
if not forward and not backward:
raise RuntimeError("Sequence does not contain any anchors")
accuracy = []
success = []
for i, reverse in [(f, False) for f in forward] + [(f, True)
for f in backward]:
name = "%s_%08d" % (sequence.name, i)
if not Trajectory.exists(results, name):
raise MissingResultsException()
if reverse:
proxy = FrameMapSequence(sequence,
list(reversed(range(0, i + 1))))
else:
proxy = FrameMapSequence(sequence,
list(range(i, sequence.length)))
trajectory = Trajectory.read(results, name)
overlaps = calculate_overlaps(
trajectory.regions(), proxy.groundtruth(),
(proxy.size) if self.burnin else None)
grace = self.grace
progress = len(proxy)
for j, overlap in enumerate(overlaps):
if overlap <= self.threshold and not proxy.groundtruth(
j).is_empty():
grace = grace - 1
if grace == 0:
progress = j + 1 - self.grace # subtract since we need actual point of the failure
break
else:
grace = self.grace
success.append((i / len(sequence), progress / len(proxy)))
accuracy.append(
(i / len(sequence), sum(overlaps[0:progress] / len(proxy))))
return success, accuracy
