def __call__(self, args, group):
print "Checking integrity..."
# read first frame to get sizes
path = Video.getframepath(0, args.location)
try:
im = Image.open(path)
except IOError:
print "Cannot read {0}".format(path)
return
width, height = im.size
print "Searching for last frame..."
# search for last frame
toplevel = max(int(x)
for x in os.listdir(args.location))
secondlevel = max(int(x)
for x in os.listdir("{0}/{1}".format(args.location, toplevel)))
maxframes = max(int(os.path.splitext(x)[0])
for x in os.listdir("{0}/{1}/{2}"
.format(args.location, toplevel, secondlevel))) + 1
print "Found {0} frames.".format(maxframes)
# can we read the last frame?
path = Video.getframepath(maxframes - 1, args.location)
try:
im = Image.open(path)
except IOError:
print "Cannot read {0}".format(path)
return
# check last frame sizes
if im.size[0] != width and im.size[1] != height:
print "First frame dimensions differs from last frame"
return
if session.query(Video).filter(Video.slug == args.slug).count():
print "Video {0} already exists!".format(args.slug)
return
if args.train_with:
if args.for_training:
print "A training video cannot require training"
return
print "Looking for training video..."
trainer = session.query(Video)
trainer = trainer.filter(Video.slug == args.train_with)
if not trainer.count():
print ("Training video {0} does not exist!"
.format(args.train_with))
return
trainer = trainer.one()
else:
trainer = None
# create video
video = Video(slug = args.slug,
location = os.path.realpath(args.location),
width = width,
height = height,
totalframes = maxframes,
skip = args.skip,
perobjectbonus = args.per_object_bonus,
completionbonus = args.completion_bonus,
trainwith = trainer,
isfortraining = args.for_training,
blowradius = args.blow_radius,
action = args.action,
pose = args.pose)
if args.for_training:
video.trainvalidator = qa.tolerable(args.for_training_overlap,
args.for_training_tolerance,
args.for_training_mistakes)
print "Training validator is {0}".format(video.trainvalidator)
session.add(video)
print "Binding labels and attributes..."
# create labels and attributes
labelcache = {}
attributecache = {}
lastlabel = None
for labeltext in args.labels:
if labeltext[0] == "~":
if lastlabel is None:
print "Cannot assign an attribute without a label!"
return
labeltext = labeltext[1:]
attribute = Attribute(text = labeltext)
session.add(attribute)
lastlabel.attributes.append(attribute)
attributecache[labeltext] = attribute
else:
label = Label(text = labeltext)
session.add(label)
video.labels.append(label)
labelcache[labeltext] = label
lastlabel = label
print "Creating symbolic link..."
symlink = "public/frames/{0}".format(video.slug)
try:
os.remove(symlink)
except:
pass
os.symlink(video.location, symlink)
print "Creating segments..."
# create shots and jobs
if args.for_training:
segment = Segment(video = video)
if args.for_training_start:
segment.start = args.for_training_start
if segment.start < 0:
segment.start = 0
else:
segment.start = 0
if args.for_training_stop:
segment.stop = args.for_training_stop
if segment.stop > video.totalframes - 1:
segment.stop = video.totalframes - 1
else:
segment.stop = video.totalframes - 1
job = Job(segment = segment, group = group, ready = False)
session.add(segment)
session.add(job)
elif args.use_frames:
with open(args.use_frames) as useframes:
for line in useframes:
ustart, ustop = line.split()
ustart, ustop = int(ustart), int(ustop)
validlength = float(ustop - ustart)
numsegments = math.ceil(validlength / args.length)
segmentlength = math.ceil(validlength / numsegments)
for start in range(ustart, ustop, int(segmentlength)):
stop = min(start + segmentlength + args.overlap + 1,
ustop)
segment = Segment(start = start,
stop = stop,
video = video)
job = Job(segment = segment, group = group)
session.add(segment)
session.add(job)
else:
startframe = args.start_frame
stopframe = args.stop_frame
if not stopframe:
stopframe = video.totalframes - 1
for start in range(startframe, stopframe, args.length):
stop = min(start + args.length + args.overlap + 1,
stopframe)
segment = Segment(start = start,
stop = stop,
video = video)
job = Job(segment = segment, group = group)
session.add(segment)
session.add(job)
if args.per_object_bonus:
group.schedules.append(
PerObjectBonus(amount = args.per_object_bonus))
if args.completion_bonus:
group.schedules.append(
CompletionBonus(amount = args.completion_bonus))
session.add(group)
if args.for_training and args.for_training_data:
print ("Loading training ground truth annotations from {0}"
.format(args.for_training_data))
with open(args.for_training_data, "r") as file:
pathcache = {}
for line in file:
(id, xtl, ytl, xbr, ybr,
frame, outside, occluded, generated,
label) = line.split(" ")
if int(generated):
continue
if id not in pathcache:
print "Imported new path {0}".format(id)
label = labelcache[label.strip()[1:-1]]
pathcache[id] = Path(job = job, label = label)
box = Box(path = pathcache[id])
box.xtl = int(xtl)
box.ytl = int(ytl)
box.xbr = int(xbr)
box.ybr = int(ybr)
box.frame = int(frame)
box.outside = int(outside)
box.occluded = int(occluded)
pathcache[id].boxes.append(box)
session.commit()
if args.for_training:
if args.for_training and args.for_training_data:
print "Video and ground truth loaded."
else:
print "Video loaded and ready for ground truth:"
print ""
print "\t{0}".format(job.offlineurl(config.localhost))
print ""
print "Visit this URL to provide training with ground truth."
else:
print "Video loaded and ready for publication."
def __call__(self, args, group):
print "Checking integrity..."
# read first frame to get sizes
path = Video.getframepath(0, args.location)
try:
im = Image.open(path)
except IOError:
print "Cannot read {0}".format(path)
return
width, height = im.size
print "Searching for last frame..."
# search for last frame
toplevel = max(int(x) for x in os.listdir(args.location))
secondlevel = max(
int(x)
for x in os.listdir("{0}/{1}".format(args.location, toplevel)))
maxframes = max(
int(os.path.splitext(x)[0])
for x in os.listdir("{0}/{1}/{2}".format(args.location, toplevel,
secondlevel))) + 1
print "Found {0} frames.".format(maxframes)
# can we read the last frame?
path = Video.getframepath(maxframes - 1, args.location)
try:
im = Image.open(path)
except IOError:
print "Cannot read {0}".format(path)
return
# check last frame sizes
if im.size[0] != width and im.size[1] != height:
print "First frame dimensions differs from last frame"
return
if session.query(Video).filter(Video.slug == args.slug).count():
print "Video {0} already exists!".format(args.slug)
return
if args.train_with:
if args.for_training:
print "A training video cannot require training"
return
print "Looking for training video..."
trainer = session.query(Video)
trainer = trainer.filter(Video.slug == args.train_with)
if not trainer.count():
print("Training video {0} does not exist!".format(
args.train_with))
return
trainer = trainer.one()
else:
trainer = None
# create video
video = Video(slug=args.slug,
location=os.path.realpath(args.location),
width=width,
height=height,
totalframes=maxframes,
skip=args.skip,
perobjectbonus=args.per_object_bonus,
completionbonus=args.completion_bonus,
trainwith=trainer,
isfortraining=args.for_training,
blowradius=args.blow_radius)
if args.for_training:
video.trainvalidator = qa.tolerable(args.for_training_overlap,
args.for_training_tolerance,
args.for_training_mistakes)
print "Training validator is {0}".format(video.trainvalidator)
session.add(video)
print "Binding labels and attributes..."
# create labels and attributes
labelcache = {}
attributecache = {}
lastlabel = None
for labeltext in args.labels:
if labeltext[0] == "~":
if lastlabel is None:
print "Cannot assign an attribute without a label!"
return
labeltext = labeltext[1:]
attribute = Attribute(text=labeltext)
session.add(attribute)
lastlabel.attributes.append(attribute)
attributecache[labeltext] = attribute
else:
label = Label(text=labeltext)
session.add(label)
video.labels.append(label)
labelcache[labeltext] = label
lastlabel = label
print "Creating symbolic link..."
symlink = "public/frames/{0}".format(video.slug)
try:
os.remove(symlink)
except:
pass
os.symlink(video.location, symlink)
print "Creating segments..."
# create shots and jobs
if args.for_training:
segment = Segment(video=video)
if args.for_training_start:
segment.start = args.for_training_start
if segment.start < 0:
segment.start = 0
else:
segment.start = 0
if args.for_training_stop:
segment.stop = args.for_training_stop
if segment.stop > video.totalframes - 1:
segment.stop = video.totalframes - 1
else:
segment.stop = video.totalframes - 1
job = Job(segment=segment, group=group, ready=False)
session.add(segment)
session.add(job)
elif args.use_frames:
with open(args.use_frames) as useframes:
for line in useframes:
ustart, ustop = line.split()
ustart, ustop = int(ustart), int(ustop)
validlength = float(ustop - ustart)
numsegments = math.ceil(validlength / args.length)
segmentlength = math.ceil(validlength / numsegments)
for start in range(ustart, ustop, int(segmentlength)):
stop = min(start + segmentlength + args.overlap + 1,
ustop)
segment = Segment(start=start, stop=stop, video=video)
job = Job(segment=segment, group=group)
session.add(segment)
session.add(job)
else:
startframe = args.start_frame
stopframe = args.stop_frame
if not stopframe:
stopframe = video.totalframes - 1
for start in range(startframe, stopframe, args.length):
stop = min(start + args.length + args.overlap + 1, stopframe)
segment = Segment(start=start, stop=stop, video=video)
job = Job(segment=segment, group=group)
session.add(segment)
session.add(job)
if args.per_object_bonus:
group.schedules.append(
PerObjectBonus(amount=args.per_object_bonus))
if args.completion_bonus:
group.schedules.append(
CompletionBonus(amount=args.completion_bonus))
session.add(group)
if args.for_training and args.for_training_data:
print("Loading training ground truth annotations from {0}".format(
args.for_training_data))
with open(args.for_training_data, "r") as file:
pathcache = {}
for line in file:
(id, xtl, ytl, xbr, ybr, frame, outside, occluded,
generated, label) = line.split(" ")
if int(generated):
continue
if id not in pathcache:
print "Imported new path {0}".format(id)
label = labelcache[label.strip()[1:-1]]
pathcache[id] = Path(job=job, label=label)
box = Box(path=pathcache[id])
box.xtl = int(xtl)
box.ytl = int(ytl)
box.xbr = int(xbr)
box.ybr = int(ybr)
box.frame = int(frame)
box.outside = int(outside)
box.occluded = int(occluded)
pathcache[id].boxes.append(box)
session.commit()
if args.for_training:
if args.for_training and args.for_training_data:
print "Video and ground truth loaded."
else:
print "Video loaded and ready for ground truth:"
print ""
print "\t{0}".format(job.offlineurl(config.localhost))
print ""
print "Visit this URL to provide training with ground truth."
else:
print "Video loaded and ready for publication."
def __call__(self, args):
video = session.query(Video).filter(Video.slug == args.slug)
if video.count() == 0:
print "Video {0} does not exist!".format(args.slug)
raise SystemExit()
video = video.one()
print "Parsing text data"
data = {}
if args.json:
data = self.getdatajson(args.labelfile)
else:
data = self.getdatatext(args.labelfile, video.totalframes)
scale = args.scale
if args.dimensions or args.original_video or args.original_frame:
print "Computing scale"
if args.original_video:
w, h = ffmpeg.extract(args.original_video).next().size
elif args.original_frame:
w, h = Image.open(args.original_frame).size
else:
w, h = args.dimensions.split("x")
w = float(w)
h = float(h)
s = float(video.width) / w
if s * h > video.height:
s = float(video.height) / h
scale = s
print "Scale = {0}".format(scale)
segmentcount = 1
for segment in video.segments:
print "Segment {0} of {1}".format(segmentcount, len(video.segments))
segmentcount += 1
for job in segment.jobs:
for boxid in data:
label = data[boxid]['label']
boxes = data[boxid]['boxes']
query = session.query(Label).filter(Label.videoid == video.id).filter(Label.text == label)
if query.count() == 0:
continue
label = query.one()
newpath = Path(label=label)
visible = False
for frame, boxdata in boxes.iteritems():
frame = int(frame)
if frame < segment.start or segment.stop <= frame or (frame % video.blowradius != 0):
continue
newbox = Box(path=newpath)
#newbox.xtl = max(boxdata['xtl'], 0)
#newbox.ytl = max(boxdata['ytl'], 0)
#newbox.xbr = max(boxdata['xbr'], 0)
#newbox.ybr = max(boxdata['ybr'], 0)
newbox.xtl = boxdata['xtl']
newbox.ytl = boxdata['ytl']
newbox.xbr = boxdata['xbr']
newbox.ybr = boxdata['ybr']
newbox.occluded = boxdata['occluded']
newbox.outside = boxdata['outside']
newbox.generated = boxdata['generated']
newbox.frame = frame
scalebox = newbox.getbox()
scalebox = scalebox.transform(scale)
newbox.xtl = scalebox.xtl
newbox.ytl = scalebox.ytl
newbox.xbr = scalebox.xbr
newbox.ybr = scalebox.ybr
if not newbox.outside:
visible = True
if visible:
job.paths.append(newpath)
session.add(job)
session.commit()
def __call__(self, args):
video = session.query(Video).filter(Video.slug == args.slug)
if video.count() == 0:
print "Video {0} does not exist!".format(args.slug)
raise SystemExit()
video = video.one()
print "Parsing text data"
data = {}
if args.json:
data = self.getdatajson(args.labelfile)
else:
data = self.getdatatext(args.labelfile, video.totalframes)
scale = args.scale
if args.dimensions or args.original_video or args.original_frame:
print "Computing scale"
if args.original_video:
w, h = ffmpeg.extract(args.original_video).next().size
elif args.original_frame:
w, h = Image.open(args.original_frame).size
else:
w, h = args.dimensions.split("x")
w = float(w)
h = float(h)
s = float(video.width) / w
if s * h > video.height:
s = float(video.height) / h
scale = s
print "Scale = {0}".format(scale)
segmentcount = 1
for segment in video.segments:
print "Segment {0} of {1}".format(segmentcount,
len(video.segments))
segmentcount += 1
for job in segment.jobs:
for boxid in data:
label = data[boxid]['label']
boxes = data[boxid]['boxes']
query = session.query(Label).filter(
Label.videoid == video.id).filter(Label.text == label)
if query.count() == 0:
continue
label = query.one()
newpath = Path(label=label)
visible = False
for frame, boxdata in boxes.iteritems():
frame = int(frame)
if frame < segment.start or segment.stop <= frame or (
frame % video.blowradius != 0):
continue
newbox = Box(path=newpath)
#newbox.xtl = max(boxdata['xtl'], 0)
#newbox.ytl = max(boxdata['ytl'], 0)
#newbox.xbr = max(boxdata['xbr'], 0)
#newbox.ybr = max(boxdata['ybr'], 0)
newbox.xtl = boxdata['xtl']
newbox.ytl = boxdata['ytl']
newbox.xbr = boxdata['xbr']
newbox.ybr = boxdata['ybr']
newbox.occluded = boxdata['occluded']
newbox.outside = boxdata['outside']
newbox.generated = boxdata['generated']
newbox.frame = frame
scalebox = newbox.getbox()
scalebox = scalebox.transform(scale)
newbox.xtl = scalebox.xtl
newbox.ytl = scalebox.ytl
newbox.xbr = scalebox.xbr
newbox.ybr = scalebox.ybr
if not newbox.outside:
visible = True
if visible:
job.paths.append(newpath)
session.add(job)
session.commit()
def __call__(self, args, group):
print "Checking integrity..."
# read first frame to get sizes
path = Video.getframepath(0, args.location)
try:
im = Image.open(path)
except IOError:
print "Cannot read {0}".format(path)
return
width, height = im.size
print "Searching for last frame..."
# search for last frame
toplevel = max(int(x) for x in os.listdir(args.location))
secondlevel = max(
int(x)
for x in os.listdir("{0}/{1}".format(args.location, toplevel)))
maxframes = max(
int(os.path.splitext(x)[0])
for x in os.listdir("{0}/{1}/{2}".format(args.location, toplevel,
secondlevel))) + 1
print "Found {0} frames.".format(maxframes)
# can we read the last frame?
path = Video.getframepath(maxframes - 1, args.location)
try:
im = Image.open(path)
except IOError:
print "Cannot read {0}".format(path)
return
# check last frame sizes
if im.size[0] != width and im.size[1] != height:
print "First frame dimensions differs from last frame"
return
if session.query(Video).filter(Video.slug == args.slug).count():
print "Video {0} already exists!".format(args.slug)
return
if args.train_with:
if args.for_training:
print "A training video cannot require training"
return
print "Looking for training video..."
trainer = session.query(Video)
trainer = trainer.filter(Video.slug == args.train_with)
if not trainer.count():
print("Training video {0} does not exist!".format(
args.train_with))
return
trainer = trainer.one()
else:
trainer = None
# create video
video = Video(slug=args.slug,
location=os.path.realpath(args.location),
width=width,
height=height,
totalframes=maxframes,
skip=args.skip,
perobjectbonus=args.per_object_bonus,
completionbonus=args.completion_bonus,
trainwith=trainer,
isfortraining=args.for_training,
blowradius=args.blow_radius)
if args.for_training:
video.trainvalidator = qa.tolerable(args.for_training_overlap,
args.for_training_tolerance,
args.for_training_mistakes)
print "Training validator is {0}".format(video.trainvalidator)
session.add(video)
print "Binding labels and attributes..."
# create labels and attributes
labelcache = {}
attributecache = {}
lastlabel = None
for labeltext in args.labels:
if labeltext[0] == "~":
if lastlabel is None:
print "Cannot assign an attribute without a label!"
return
labeltext = labeltext[1:]
attribute = Attribute(text=labeltext)
session.add(attribute)
lastlabel.attributes.append(attribute)
attributecache[labeltext] = attribute
else:
label = Label(text=labeltext)
session.add(label)
video.labels.append(label)
labelcache[labeltext] = label
lastlabel = label
print "Creating symbolic link..."
symlink = "public/frames/{0}".format(video.slug)
try:
os.remove(symlink)
except:
pass
os.symlink(video.location, symlink)
print "Creating segments..."
# create shots and jobs
job_list = []
segment_list = []
if args.for_training:
segment = Segment(video=video)
if args.for_training_start:
segment.start = args.for_training_start
if segment.start < 0:
segment.start = 0
else:
segment.start = 0
if args.for_training_stop:
segment.stop = args.for_training_stop
if segment.stop > video.totalframes - 1:
segment.stop = video.totalframes - 1
else:
segment.stop = video.totalframes - 1
job = Job(segment=segment, group=group, ready=False)
job_list.append(job)
segment_list.append(segment)
session.add(segment)
session.add(job)
elif args.use_frames:
with open(args.use_frames) as useframes:
for line in useframes:
ustart, ustop = line.split()
ustart, ustop = int(ustart), int(ustop)
validlength = float(ustop - ustart)
numsegments = math.ceil(validlength / args.length)
segmentlength = math.ceil(validlength / numsegments)
for start in range(ustart, ustop, int(segmentlength)):
stop = min(start + segmentlength + args.overlap + 1,
ustop)
segment = Segment(start=start, stop=stop, video=video)
job = Job(segment=segment, group=group)
job_list.append(job)
segment_list.append(segment)
session.add(segment)
session.add(job)
else:
startframe = args.start_frame
stopframe = args.stop_frame
if not stopframe:
stopframe = video.totalframes - 1
for start in range(startframe, stopframe, args.length):
stop = min(start + args.length + args.overlap + 1, stopframe)
segment = Segment(start=start, stop=stop, video=video)
job = Job(segment=segment, group=group)
job_list.append(job)
segment_list.append(segment)
session.add(segment)
session.add(job)
if args.per_object_bonus:
group.schedules.append(
PerObjectBonus(amount=args.per_object_bonus))
if args.completion_bonus:
group.schedules.append(
CompletionBonus(amount=args.completion_bonus))
session.add(group)
if args.for_training and args.for_training_data:
print("Loading training ground truth annotations from {0}".format(
args.for_training_data))
with open(args.for_training_data, "r") as file:
pathcache = {}
for line in file:
(id, xtl, ytl, xbr, ybr, frame, outside, occluded,
generated, label) = line.split(" ")
if int(generated):
continue
if id not in pathcache:
print "Imported new path {0}".format(id)
label = labelcache[label.strip()[1:-1]]
pathcache[id] = Path(job=job, label=label)
box = Box(path=pathcache[id])
box.xtl = int(xtl)
box.ytl = int(ytl)
box.xbr = int(xbr)
box.ybr = int(ybr)
box.frame = int(frame)
box.outside = int(outside)
box.occluded = int(occluded)
pathcache[id].boxes.append(box)
session.commit()
# Save the annotated file in the database
if args.annot is not None:
with open(args.annot, 'r') as annot_file:
annotated_tracks = json.load(annot_file)
# Scale annotations if annot-video-height is given as argument
annot_scalar = 1.0
if args.annot_video_height is not None:
annot_scalar = video.height / float(
args.annot_video_height) * 1.5
print('Scale factor: {}'.format(annot_scalar))
# Scale bboxes and convert labels
converted_tracks = []
for a_labels, a_tracks, a_attribs in annotated_tracks:
scaled_a_tracks = {}
keep_tracks = a_tracks.keys()
# Blow-radius for annotations
if args.annot_blow_radius > 0:
keep_tracks = sorted(map(int, keep_tracks))
prev_frame_id = keep_tracks[0]
blown_tracks = [keep_tracks[0]]
for frame_id in keep_tracks:
if frame_id > prev_frame_id + args.annot_blow_radius:
prev_frame_id = frame_id
blown_tracks.append(frame_id)
keep_tracks = map(unicode, blown_tracks)
# Conversion
for track_id, track_data in a_tracks.iteritems():
if track_id in keep_tracks:
scaled_track = [
x * annot_scalar for x in track_data[:4]
]
scaled_track.extend(track_data[4:])
scaled_a_tracks[track_id] = scaled_track
converted_tracks.append(
[labelcache[a_labels].id, scaled_a_tracks, a_attribs])
for j, s in zip(job_list, segment_list):
job_data = []
for a_labels, a_tracks, a_attribs in converted_tracks:
# Sort by track number
sorted_a_tracks = natural_sort(a_tracks.keys())
sorted_a_attribs = natural_sort(a_attribs.keys())
job_tracks = {}
job_attribs = {k.id: {} for k in attributecache.values()}
for track_id in sorted_a_tracks:
track_data = a_tracks[track_id]
track_id = int(track_id)
if s.start <= track_id <= s.stop:
job_tracks[track_id] = track_data
if track_id > s.stop:
break
for track_id in sorted_a_attribs:
attrib_data = a_attribs[track_id]
track_id = int(track_id)
if s.start <= track_id <= s.stop:
for k in job_attribs.keys():
job_attribs[k][track_id] = 0
job_attribs[
attributecache[attrib_data].id][track_id] = 1
if track_id > s.stop:
break
job_data.append([a_labels, job_tracks, job_attribs])
savejob(j.id, job_data)
if args.for_training:
if args.for_training and args.for_training_data:
print "Video and ground truth loaded."
else:
print "Video loaded and ready for ground truth:"
print ""
print "\t{0}".format(job.offlineurl(config.localhost))
print ""
print "Visit this URL to provide training with ground truth."
else:
print "Video loaded and ready for publication."