class Experiment(Attributee):
realtime = Nested(RealtimeConfig, default=None)
noise = Nested(NoiseConfig, default=None)
inject = Nested(InjectConfig, default=None)
transformers = List(Object(transformer_resolver), default=[])
analyses = List(Object(analysis_resolver), default=[])
def __init__(self, _identifier: str, _storage: "LocalStorage", **kwargs):
self._identifier = _identifier
self._storage = _storage
super().__init__(**kwargs)
# TODO: validate analysis names
@property
def identifier(self) -> str:
return self._identifier
@property
def storage(self) -> "Storage":
return self._storage
def _get_initialization(self, sequence: "Sequence", index: int):
return sequence.groundtruth(index)
def _get_runtime(self, tracker: "Tracker", sequence: "Sequence"):
if not self.realtime is None:
grace = to_number(self.realtime.grace, min_n=0)
fps = to_number(self.realtime.fps, min_n=0, conversion=float)
interval = 1 / float(sequence.metadata("fps", fps))
runtime = RealtimeTrackerRuntime(tracker.runtime(), grace,
interval)
else:
runtime = tracker.runtime()
return runtime
@abstractmethod
def execute(self,
tracker: "Tracker",
sequence: "Sequence",
force: bool = False,
callback: typing.Callable = None):
raise NotImplementedError
@abstractmethod
def scan(self, tracker: "Tracker", sequence: "Sequence"):
raise NotImplementedError
def results(self, tracker: "Tracker", sequence: "Sequence") -> "Results":
return self._storage.results(tracker, self, sequence)
def log(self, identifier: str):
return self._storage.substorage("logs").write(
"{}_{:%Y-%m-%dT%H-%M-%S.%f%z}.log".format(identifier,
datetime.now()))
def transform(self, sequence: "Sequence"):
for transformer in self.transformers:
sequence = transformer(sequence)
return sequence
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 ReportConfiguration(Attributee):
style = Nested(StyleManager)
sort = Nested(TrackerSorter)
generators = List(Object(subclass=Generator), default=[])
class AttributeMultiStart(SequenceAveragingAnalysis):
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)
tags = List(String())
@property
def name(self):
return "AR per-attribute analysis"
def describe(self):
accuracy = [
Measure("Accuracy: " + t,
"A " + t,
minimal=0,
maximal=1,
direction=Sorting.DESCENDING) for t in self.tags
]
robustness = [
Measure("Robutsness" + t,
"R " + t,
minimal=0,
maximal=1,
direction=Sorting.DESCENDING) for t in self.tags
]
length = [None] * len(self.tags)
return tuple(
functools.reduce(
operator.add,
[[a, r, n] for a, r, n in zip(accuracy, robustness, length)]))
def compatible(self, experiment: Experiment):
return isinstance(experiment, MultiStartExperiment)
def collapse(self, tracker: Tracker, sequences: typing.List[Sequence],
results: typing.List[tuple]):
accuracy = Counter()
robustness = Counter()
attribute_total = Counter()
for seq_acc, seq_rob, seq_attr_count in results:
for t in seq_attr_count:
accuracy[t] += (seq_acc[t]
if t in seq_acc else 0) * seq_attr_count[t]
robustness[t] += seq_rob * seq_attr_count[t]
attribute_total[t] += seq_attr_count[t]
accuracy = [accuracy[t] / attribute_total[t] for t in self.tags]
robustness = [robustness[t] / attribute_total[t] for t in self.tags]
length = [attribute_total[t] for t in self.tags]
return tuple(
functools.reduce(
operator.add,
[[a, r, n] for a, r, n in zip(accuracy, robustness, length)]))
def subcompute(self, experiment: Experiment, tracker: Tracker,
sequence: Sequence):
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")
accuracy_ = Counter()
tags_count_ = Counter()
robustness_ = 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
for j in range(progress):
overlap = overlaps[j]
tags = proxy.tags(j)
if len(tags) == 0:
tags = ['empty']
for t in tags:
accuracy_[t] += overlap
tags_count_[t] += 1
robustness_ += progress
total_ += len(proxy)
seq_robustness = robustness_ / total_
seq_accuracy = {}
for t in accuracy_:
seq_accuracy[t] = accuracy_[t] / tags_count_[t]
# calculate weights for each attribute
attribute_counter = Counter()
for frame_idx in range(len(sequence)):
tags = sequence.tags(frame_idx)
if len(tags) == 0:
tags = ['empty']
for t in tags:
attribute_counter[t] += 1
return seq_accuracy, seq_robustness, attribute_counter
class AttributeDifficultyLevelMultiStart(SequenceAveragingAnalysis):
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)
fail_interval = Integer(default=30, val_min=1)
tags = List(String())
@property
def name(self):
return "Attribute difficulty"
def describe(self):
return tuple([
Measure(t, t, minimal=0, maximal=1, direction=Sorting.DESCENDING)
for t in self.tags
] + [None] * len(self.tags))
def compatible(self, experiment: Experiment):
return isinstance(experiment, MultiStartExperiment)
def collapse(self, tracker: Tracker, sequences: typing.List[Sequence],
results: typing.List[tuple]):
attribute_difficulty = Counter()
attribute_counter = Counter()
for seq_tags_not_failed, seq_tags_count, seq_attr_count in results:
for tag in seq_tags_count:
if tag in seq_tags_not_failed:
seq_attr_difficulty = seq_tags_not_failed[
tag] / seq_tags_count[tag]
else:
seq_attr_difficulty = 0
attribute_difficulty[
tag] += seq_attr_difficulty * seq_attr_count[tag]
attribute_counter[tag] += seq_attr_count[tag]
return tuple([
attribute_difficulty[tag] / attribute_counter[tag]
for tag in self.tags
] + [attribute_counter[tag] for tag in self.tags])
def subcompute(self, experiment: Experiment, tracker: Tracker,
sequence: Sequence):
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")
tags_count = Counter()
tags_not_failed = Counter()
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
for j in range(progress):
tags = proxy.tags(j)
if len(tags) == 0:
tags = ['empty']
for t in tags:
tags_count[t] += 1
if progress == len(
proxy) or j < progress - self.fail_interval:
tags_not_failed[t] += 1
attribute_counter = Counter()
for frame_idx in range(len(sequence)):
tags = sequence.tags(frame_idx)
if len(tags) == 0:
tags = ['empty']
for t in tags:
attribute_counter[t] += 1
return tags_not_failed, tags_count, attribute_counter
class Workspace(Attributee):
registry = List(
String(transformer=lambda x, ctx: normalize_path(
x, ctx["parent"].directory)))
stack = StackLoader()
sequences = String(default="sequences")
report = Nested(ReportConfiguration)
@staticmethod
def initialize(directory, config=None, download=False):
config_file = os.path.join(directory, "config.yaml")
if os.path.isfile(config_file):
raise WorkspaceException("Workspace already initialized")
os.makedirs(directory, exist_ok=True)
with open(config_file, 'w') as fp:
yaml.dump(config if config is not None else dict(), fp)
os.makedirs(os.path.join(directory, "sequences"), exist_ok=True)
os.makedirs(os.path.join(directory, "results"), exist_ok=True)
if not os.path.isfile(os.path.join(directory, "trackers.ini")):
open(os.path.join(directory, "trackers.ini"), 'w').close()
download = False
if download:
# Try do retrieve dataset from stack and download it
stack_file = resolve_stack(config["stack"], directory)
dataset_directory = normalize_path(
config.get("sequences", "sequences"), directory)
if stack_file is None:
return
dataset = None
with open(stack_file, 'r') as fp:
stack_metadata = yaml.load(fp, Loader=yaml.BaseLoader)
dataset = stack_metadata["dataset"]
if dataset:
Workspace.download_dataset(dataset, dataset_directory)
@staticmethod
def download_dataset(dataset, directory):
if os.path.exists(os.path.join(directory, "list.txt")):
return False
from vot.dataset import download_dataset
download_dataset(dataset, directory)
logger.info("Download completed")
@staticmethod
def load(directory):
directory = normalize_path(directory)
config_file = os.path.join(directory, "config.yaml")
if not os.path.isfile(config_file):
raise WorkspaceException("Workspace not initialized")
with open(config_file, 'r') as fp:
config = yaml.load(fp, Loader=yaml.BaseLoader)
return Workspace(directory, **config)
def __init__(self, directory, **kwargs):
self._directory = directory
self._storage = LocalStorage(
directory) if directory is not None else VoidStorage()
super().__init__(**kwargs)
dataset_directory = normalize_path(self.sequences, directory)
if not self.stack.dataset is None:
Workspace.download_dataset(self.stack.dataset, dataset_directory)
self._dataset = VOTDataset(dataset_directory)
@property
def directory(self) -> str:
return self._directory
@property
def dataset(self) -> Dataset:
return self._dataset
@property
def storage(self) -> LocalStorage:
return self._storage
def cache(self, identifier) -> LocalStorage:
if not isinstance(identifier, str):
identifier = class_fullname(identifier)
return self._storage.substorage("cache").substorage(identifier)
def list_results(self, registry: "Registry"):
references = self._storage.substorage("results").folders()
return registry.resolve(*references)