def readpaths(tracks):
paths = []
logger.debug("Reading {0} total tracks".format(len(tracks)))
for label, track, attributes in tracks:
path = Path()
path.label = session.query(Label).get(label)
logger.debug("Received a {0} track".format(path.label.text))
for frame, userbox in track.items():
box = Box(path = path)
box.xtl = int(userbox[0])
box.ytl = int(userbox[1])
box.xbr = int(userbox[2])
box.ybr = int(userbox[3])
box.occluded = int(userbox[4])
box.outside = int(userbox[5])
box.frame = int(frame)
logger.debug("Received box {0}".format(str(box.getbox())))
for attributeid, timeline in attributes.items():
attribute = session.query(Attribute).get(attributeid)
for frame, value in timeline.items():
aa = AttributeAnnotation()
aa.attribute = attribute
aa.frame = frame
aa.value = value
path.attributes.append(aa)
paths.append(path)
return paths
def __call__(self, args):
video = session.query(Video).filter(Video.slug == args.slug)
if not video.count():
print "Video {0} does not exist!".format(args.slug)
return
video = video.one()
query = session.query(Path)
query = query.join(Job)
query = query.join(Segment)
query = query.filter(Segment.video == video)
numpaths = query.count()
if numpaths and not args.force:
print ("Video has {0} paths. Use --force to delete."
.format(numpaths))
return
for segment in video.segments:
for job in segment.jobs:
if job.published and not job.completed:
hitid = job.disable()
print "Disabled {0}".format(hitid)
session.delete(video)
session.commit()
print "Deleted video and associated data."
def __call__(self, args):
query = session.query(HIT)
query = query.filter(HIT.useful == True)
if args.hit:
query = query.filter(HIT.hitid == args.id)
else:
worker = session.query(Worker).get(args.id)
if not worker:
print "Worker \"{0}\" not found".format(args.id)
return
if not args.no_block:
worker.block("HIT was invalid.")
print "Blocked worker \"{0}\"".format(args.id)
session.add(worker)
query = query.filter(HIT.workerid == args.id)
for hit in query:
replacement = hit.invalidate()
session.add(hit)
print "Invalidated {0}".format(hit.hitid)
if replacement:
session.add(replacement)
if not args.no_publish:
session.commit()
replacement.publish()
session.add(replacement)
print "Respawned with {0}".format(replacement.hitid)
session.commit()
def __call__(self, args):
videos = session.query(Video)
if args.training:
videos = videos.filter(Video.isfortraining == True)
else:
videos = videos.filter(Video.isfortraining == False)
if args.worker:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.workerid == args.worker)
elif args.published:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.published == True)
elif args.completed:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.completed == True)
elif args.incomplete:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.completed == False)
if args.count:
print videos.count()
else:
newvideos = dict()
print "Identifier", '-------', 'jobid', '------', 'timeonserver', ' ------------------- HitId', ' ------------------- AssignmentId', ' --------------- WorkerId'
for video in videos.distinct():
# print video.slug
# Print videos sorted by time
test = session.query(Job).filter(Job.useful == True)
test = test.join(Segment).filter(Segment.videoid == video.id)
if test.count() != 1:
print "Error: ", test.count()
break
for t in test:
l = list()
if t.timeonserver is None:
l.append(datetime.datetime.now())
else:
l.append(t.timeonserver)
l.append(t)
newvideos[video.slug] = l
sorted_list = [x for x in newvideos.iteritems()]
sorted_list.sort(key=lambda x: x[1][0]) # sort by key
for k in sorted_list:
print k[0], "---", k[1][1].id, "---", k[1][0], "---", k[1][1].hitid, "---", k[1][1].assignmentid, "---", k[1][1].workerid
def upload(id, environ):
job = session.query(Job).get(id)
data = environ["wsgi.input"]
# read meta data first
header = data.readline().strip() + "--"
while True:
chunk = data.readline()
if chunk.strip() == "":
break
key, value = chunk.split(": ", 1)
if key == "Content-Type":
job.mimetype = value.strip()
if key == "Content-Disposition":
for item in value.split("; "):
itemdata = item.split("=", 1)
if len(itemdata) == 2 and itemdata[0] == "filename":
job.filename = itemdata[1].strip()[1:-1]
session.commit()
# now read the file data, looking for the terminating sequence
out = open(job.storepath, "wb")
while True:
chunk = data.readline(1024 * 1024)
if chunk.strip() == header:
break
out.write(chunk)
out.close()
return ["<script>parent.uploaded();</script>"]
def validatejob(id, data):
job = session.query(Job).get(id)
paths = readpaths(data["tracks"])
predicates = readpredicates(data["predicates"])
return (job.trainingjob.validator(paths, job.trainingjob.paths) and
job.trainingjob.validator(predicates, job.trainingjob.predicates))
def savejob(id, data):
job = session.query(Job).get(id)
for path in job.paths:
session.delete(path)
for pi in job.predicate_instances:
for pa in pi.predicate_annotations:
session.delete(pa)
session.delete(pi)
for s in job.sentences:
for sa in s.annotations:
session.delete(sa)
session.delete(s)
session.commit()
paths = readpaths(data["tracks"])
for path in paths:
job.paths.append(path)
for pi in readpredicates(data["predicates"], paths):
job.predicate_instances.append(pi)
for s in readsentences(data['sentences']):
job.sentences.append(s)
session.add(job)
session.commit()
def trackforward(id, frame, tracker, trackid, tracks):
frame = int(frame)
trackid = int(trackid)
job = session.query(Job).get(id)
segment = job.segment
video = segment.video
paths = [path for path in readpaths(tracks) if path is not None]
paths = trackutils.totrackpaths(paths)
logger.info("Job Id: {0}".format(id))
logger.info("Algorithm: {0}".format(tracker))
start = frame
stop = segment.stop
outpath = tracking.api.online(tracker, start, stop, video.location, trackid, paths)
path = trackutils.fromtrackpath(outpath, job, start, stop)
attrs = [(x.attributeid, x.frame, x.value) for x in path.attributes]
logger.info("Path: {0}".format(path))
return {
"label": 0,
"boxes": [tuple(x) for x in path.getboxes()],
"attributes": attrs
}
def videodump(slug, outputtype, groundplane, fields=None):
logger.debug(os.getcwd())
query = session.query(Video).filter(Video.slug == slug)
if query.count() != 1:
raise ValueError("Invalid video slug")
video = query.one()
#mergemethod = merge.userid
groundplane = (groundplane == 1)
mergemethod = merge.getpercentoverlap(groundplane)
if fields is None:
if groundplane:
fields = dumptools.GROUND_PLANE_FORMAT
else:
fields = dumptools.DEFAULT_FORMAT
fields = fields.split()
data = dumptools.getdata(video, True, mergemethod, 0.5, None, groundplane)
outfile = tempfile.TemporaryFile()
if outputtype == "json":
dumptools.dumpjson(outfile, data, groundplane, fields)
elif outputtype == "xml":
dumptools.dumpxml(outfile, data, groundplane, fields)
else:
dumptools.dumptext(outfile, data, groundplane, fields)
outfile.seek(0)
text = outfile.readlines()
outfile.close()
return text
def getallvideos():
query = session.query(Video)
videos = []
for video in query:
newvideo = {
"slug": video.slug,
"segments": [],
}
for segment in video.segments:
newsegment = {
"start": segment.start,
"stop":segment.stop,
"jobs":[],
}
for job in segment.jobs:
newsegment["jobs"].append({
"url": job.offlineurl(config.localhost),
"numobjects": len(job.paths),
"numdone": len([path for path in job.paths if path.done]),
})
newvideo["segments"].append(newsegment)
videos.append(newvideo)
return videos
def __call__(self, args):
videos = session.query(Video)
if args.training:
videos = videos.filter(Video.isfortraining == True)
else:
videos = videos.filter(Video.isfortraining == False)
if args.worker:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.workerid == args.worker)
elif args.published:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.published == True)
elif args.completed:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.completed == True)
if args.count:
print videos.count()
else:
for video in videos.distinct():
print "{0:<25}".format(video.slug),
if args.stats:
print "{0:>3}/{1:<8}".format(video.numcompleted, video.numjobs),
print "${0:<15.2f}".format(video.cost),
print ""
def __call__(self, args):
videos = session.query(Video)
if args.training:
videos = videos.filter(Video.isfortraining == True)
else:
videos = videos.filter(Video.isfortraining == False)
if args.worker:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.workerid == args.worker)
elif args.published:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.published == True)
elif args.completed:
videos = videos.join(Segment)
videos = videos.join(Job)
videos = videos.filter(Job.completed == True)
if args.count:
print videos.count()
else:
for video in videos.distinct():
print video.slug
def savetopview(slug, image, environ):
logger.info("Saving topview image")
query = session.query(Video).filter(Video.slug == slug)
if query.count() != 1:
raise ValueError("Invalid video slug")
video = query[0]
savedir = video.homographylocation
if savedir is None:
savedir = makehomographydir(video)
savelocation = os.path.join(savedir, "topview.jpg")
tempformfile = tempfile.TemporaryFile()
tempformfile.write(image)
tempformfile.seek(0)
form = cgi.FieldStorage(fp=tempformfile, environ=environ, keep_blank_values=True)
outfile = open(savelocation, "w+b")
shutil.copyfileobj(form['photo'].file, outfile)
tempformfile.close()
outfile.close()
newimage = cv2.imread(savelocation)
scale = 1
if newimage.shape[1] > video.width:
scale = float(video.width) / float(newimage.shape[1])
newimage = cv2.resize(newimage, (0, 0), None, scale, scale)
if newimage.shape[0] > video.height:
scale = float(video.height) / float(newimage.shape[0])
newimage = cv2.resize(newimage, (0, 0), None, scale, scale)
cv2.imwrite(savelocation, newimage)
def __call__(self, args):
jobs = session.query(Job)
jobs = jobs.join(Segment).join(Video)
if args.id:
jobs = jobs.filter(Video.slug == args.id)
if args.frame is not None:
jobs = jobs.filter(Segment.start <= args.frame)
jobs = jobs.filter(Segment.stop >= args.frame)
if args.hitid:
jobs = jobs.filter(Job.hitid == args.hitid)
if args.workerid:
jobs = jobs.filter(Job.workerid == args.workerid)
jobs = jobs.filter(turkic.models.HIT.useful == True)
if jobs.count() > 0:
for job in jobs:
if args.ids:
if job.published:
print job.hitid,
if job.completed:
print job.assignmentid,
print job.workerid,
print ""
else:
print "(not published)"
else:
print job.offlineurl(config.localhost)
else:
print "No jobs matching this criteria."
def __call__(self, args):
query = session.query(Job).filter(turkic.models.HIT.completed == True)
for job in query:
print job.storepath
print job.mimetype
for activity in job.activities:
print activity.text
print ""
def getboxesforjob(id):
job = session.query(Job).get(id)
result = []
for path in job.paths:
attrs = [(x.attributeid, x.frame, x.value) for x in path.attributes]
result.append({"label": path.labelid,
"boxes": [tuple(x) for x in path.getboxes()],
"attributes": attrs})
return result
def markcomplete(hitid, assignmentid, workerid):
"""
Marks a job as complete. Usually this is called right before the
MTurk form is submitted.
"""
hit = session.query(models.HIT).filter(models.HIT.hitid == hitid).one()
hit.markcompleted(workerid, assignmentid)
session.add(hit)
session.commit()
def getjob(id, verified):
job = session.query(Job).get(id)
logger.debug("Found job {0}".format(job.id))
if int(verified) and job.segment.video.trainwith:
# swap segment with the training segment
training = True
segment = job.segment.video.trainwith.segments[0]
logger.debug("Swapping actual segment with training segment")
else:
training = False
segment = job.segment
video = segment.video
labels = dict((l.id, l.text) for l in video.labels)
attributes = {}
for label in video.labels:
attributes[label.id] = dict((a.id, a.text) for a in label.attributes)
roles = dict((r.id, r.text) for r in session.query(Role))
predicates = dict((p.id, p.text) for p in session.query(Predicate))
logger.debug("Giving user frames {1} to {2} of {0}".format(video.slug,
segment.start,
segment.stop))
return {"start": segment.start,
"stop": segment.stop,
"slug": video.slug,
"width": video.width,
"height": video.height,
"skip": video.skip,
"perobject": video.perobjectbonus,
"completion": video.completionbonus,
"blowradius": video.blowradius,
"jobid": job.id,
"training": int(training),
"labels": labels,
"attributes": attributes,
"roles": roles,
"predicates": predicates}
def getsentenceannotationsforjob(id):
job = session.query(Job).get(id)
results = []
for s in job.sentences:
annotations = [(a.frame, a.value) for a in s.annotations]
results.append({
'sentence': s.text,
'annotations': annotations,
})
return results
def savedonationstatus(hitid, donation):
"""
Saves the donation statistics
"""
hit = session.query(models.HIT).filter(models.HIT.hitid == hitid).one()
hit.opt2donate = float(donation)
hit.opt2donate = min(max(hit.opt2donate, 0), 1)
session.add(hit)
session.commit()
def readpath(label, userid, done, track, attributes):
path = Path()
path.label = session.query(Label).get(label)
path.done = int(done)
path.userid = int(userid)
logger.debug("Received a {0} track".format(path.label.text))
visible = False
for frame, userbox in track.items():
box = Box(path = path)
#box.xtl = max(int(userbox[0]), 0)
#box.ytl = max(int(userbox[1]), 0)
#box.xbr = max(int(userbox[2]), 0)
#box.ybr = max(int(userbox[3]), 0)
box.xtl = int(userbox[0])
box.ytl = int(userbox[1])
box.xbr = int(userbox[2])
box.ybr = int(userbox[3])
box.occluded = int(userbox[4])
box.outside = int(userbox[5])
box.generated = int(userbox[6])
box.frame = int(frame)
if not box.outside:
visible = True
logger.debug("Received box {0}".format(str(box.getbox())))
if not visible:
logger.warning("Received empty path! Skipping")
return
for attributeid, timeline in attributes.items():
attribute = session.query(Attribute).get(attributeid)
for frame, value in timeline.items():
aa = AttributeAnnotation()
aa.attribute = attribute
aa.frame = frame
aa.value = value
path.attributes.append(aa)
return path
def savejob(id, tracks):
job = session.query(Job).get(id)
for path in job.paths:
session.delete(path)
session.commit()
for path in readpaths(tracks):
job.paths.append(path)
session.add(job)
session.commit()
def saveeventlog(hitid, events):
"""
Records the event log to database.
"""
hit = session.query(models.HIT).filter(models.HIT.hitid == hitid).one()
for timestamp, domain, message in events:
timestamp = datetime.fromtimestamp(int(timestamp) / 1000)
event = EventLog(hit=hit, domain=domain, message=message, timestamp=timestamp)
session.add(event)
session.commit()
def trackfull(id, tracker):
job = session.query(Job).get(id)
segment = job.segment
video = segment.video
tracks = tracking.runfulltracker(tracker, segment.start, segment.stop, video.location)
allpaths = [convert_track_to_path(segment.start, track, job) for track in tracks]
allattrs = [[(x.attributeid, x.frame, x.value) for x in path.attributes] for path in allpaths]
return [{
"label": 0,
"boxes":[tuple(x) for x in path.getboxes()],
"attributes":attrs,
} for path, attrs in zip(allpaths, allattrs)]
def readpredicates(predicates, paths):
predicateInstances = []
logger.debug("Reading {0} total predicate instances".format(len(predicates)))
for p in predicates:
pi = PredicateInstance()
pi.predicate = session.query(Predicate).get(int(p['predicate']))
for pathid in p['annotations'].keys():
path = paths[int(pathid)]
for frame, roleid, value in p['annotations'][pathid]:
pa = PredicateAnnotation()
pa.predicateinstance = pi
pa.path = path
pa.role = session.query(Role).get(roleid)
pa.frame = frame
pa.value = value
predicateInstances.append(pi)
return predicateInstances
def respawnjob(id):
job = session.query(Job).get(id)
replacement = job.markastraining()
job.worker.verified = True
session.add(job)
session.add(replacement)
session.commit()
replacement.publish()
session.add(replacement)
session.commit()
def serverstatus(self, session):
available = session.query(HIT).filter(HIT.ready == True).count()
published = session.query(HIT).filter(HIT.published == True).count()
completed = session.query(HIT).filter(HIT.completed == True).count()
compensated = session.query(HIT).filter(HIT.compensated == True).count()
remaining = published - completed
print "Status:"
print " Available: {0}".format(available)
print " Published: {0}".format(published)
print " Completed: {0}".format(completed)
print " Compensated: {0}".format(compensated)
print " Remaining: {0}".format(remaining)
print ""
if remaining > 0:
print "Server is ONLINE and accepting work!"
else:
if compensated == completed:
print "Server is offline."
else:
print "Server is offline, but some workers are not compensated."
def savehomography(slug, homography):
query = session.query(Video).filter(Video.slug == slug)
if query.count() != 1:
raise ValueError("Invalid video slug")
video = query[0]
savedir = video.homographylocation
if savedir is None:
savedir = makehomographydir(video)
savelocation = os.path.join(savedir, "homography.npy")
np.save(savelocation, np.array(homography))
session.add(video)
session.commit()
def getdata(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()
mergemethod = merge.getpercentoverlap(args.groundplane)
if args.merge_method == "id":
mergemethod = merge.userid
workers = None
if args.worker:
workers = set(args.worker)
return video, dumptools.getdata(video, args.merge, mergemethod, args.merge_threshold,
workers, args.groundplane)
def resetdatabase(self):
try:
hits = session.query(HIT)
hits = hits.filter(HIT.published == True)
hits = hits.filter(HIT.completed == False)
for hit in hits:
print "Disabled HIT {0}".format(hit.hitid)
hit.disable()
except:
print "Failed disabling online HITs. Disable manually with:"
print "\timport config, turkic.api"
print "\tturkic.api.server.purge()"
database.reinstall()
print "Database reset!"
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 validatejob(id, tracks):
job = session.query(Job).get(id)
paths = readpaths(tracks)
return job.trainingjob.validator(paths, job.trainingjob.paths)
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):
try:
os.makedirs(args.directory)
except:
pass
since = None
if args.since:
since = parsedatetime.Calendar().parse(args.since)
since = time.mktime(since[0])
since = datetime.datetime.fromtimestamp(since)
if args.labels:
font = ImageFont.truetype("arial.ttf", 14)
else:
font = None
workers = session.query(turkic.models.Worker)
for worker in workers:
print "Sampling worker {0}".format(worker.id)
jobs = session.query(Job)
jobs = jobs.filter(Job.worker == worker)
jobs = jobs.join(Segment)
jobs = jobs.join(Video)
jobs = jobs.filter(Video.isfortraining == False)
if since:
jobs = jobs.filter(turkic.models.HIT.timeonserver >= since)
jobs = jobs.order_by(sqlalchemy.func.rand())
jobs = jobs.limit(args.number)
for job in jobs:
print "Visualizing HIT {0}".format(job.hitid)
paths = [
x.getboxes(interpolate=True, bind=True, label=True)
for x in job.paths
]
if args.frames > job.segment.stop - job.segment.start:
frames = range(job.segment.start, job.segment.stop + 1)
else:
frames = random.sample(
xrange(job.segment.start, job.segment.stop + 1),
args.frames)
size = math.sqrt(len(frames))
video = job.segment.video
bannersize = (video.width * int(math.floor(size)),
video.height * int(math.ceil(size)))
image = Image.new(video[0].mode, bannersize)
size = int(math.floor(size))
offset = (0, 0)
horcount = 0
paths = vision.visualize.highlight_paths(video,
paths,
font=font)
for frame, framenum in paths:
if framenum in frames:
image.paste(frame, offset)
horcount += 1
if horcount >= size:
offset = (0, offset[1] + video.height)
horcount = 0
else:
offset = (offset[0] + video.width, offset[1])
image.save("{0}/{1}-{2}.jpg".format(args.directory, worker.id,
job.hitid))
def __call__(self, args):
hits = session.query(HIT).filter(HIT.donatedamount > 0)
for hit in hits:
print hit.workerid, hit.timeonserver, hit.donatedamount
def getjob(id):
# the database is called session
job = session.query(Job).get(id)
activities = [x.text for x in job.activities]
return {"activities": activities}
def addtracks(
id,
base_path='/media/leo/24DDF7874B2D4C94/FLASH_ALPHA',
fname="/media/leo/24DDF7874B2D4C94/FLASH_ALPHA/401/test/401/mse0.1_track0.01_output.txt",
resize=True):
'''
Function for adding pre-existing tracks to the database
Inputs
:fname: string, path to track file
Returns
None
'''
job = session.query(Job).get(id)
logger.debug(job)
labels = session.query(Label).all()
slug = session.query(Video).get(id)
code = slug[0:3]
path = os.path.join(base_path, slug, "test", slug)
for f in os.listdir(path):
if "output" in f:
fname = f
fname = os.path.join(path, fname)
# Print all labels
print("There are %d labels in the track" % (len(labels)))
for l in labels:
print("Label id %d corresponds to %s" % (int(l.id), l.text))
# Add labels
face_label = Label(text="Face")
tv_label = Label(text="TV")
#session.add(face_label)
#session.add(tv_label)
place = Attribute(text="Placeholder")
gaze = Attribute(text="Gaze")
no_gaze = Attribute(text="No-Gaze")
uncertain = Attribute(text="Uncertain")
out = Attribute(text="Out-of-Frame")
attribs = [place, gaze, no_gaze, uncertain, out]
for a in attribs:
face_label.attributes.append(a)
tv_label.attributes.append(Attribute(text="On"))
if job.paths == None or len(job.paths) == 0:
print("Adding tracks from file %s" % (fname))
else:
print("Tracks already exist for this segment, no need to load!")
print(len(job.paths))
print([j.id for j in job.paths])
for path in job.paths:
session.delete(path)
session.commit()
#return
info = np.loadtxt(fname, dtype=int)
"""
info[:,0] ==> frame number
info[:,1] ==> left
info[:,2] ==> top
info[:,3] ==> right
info[:,4] ==> bottom
info[:,5] ==> faceID
info[:,6] ==> pause flag
"""
"""
Even with an actual label, it still says the label is none in readpath
"""
face_label = session.query(Label).all()[0]
print(face_label.text)
print(face_label.id)
num_faces = np.max(info[:, 5])
scale = 1
if resize:
scale = .375
print("There are %d unique faces detected." % (num_faces))
thresh = 10
for newID in range(num_faces):
print(newID)
if newID == thresh:
break
path = Path(job=job, label=session.query(Label).get(82))
# uncomment below only if necessary
#path.label = 100 # label is some constant integer value
#path.label = Label(text = "Face")
path.userid = newID # userid is the label ID
lines = info[np.where(info[:, 5] == newID)[0], :]
boxes = []
# need to scale within range 1920x1080 -> 720x405
#
# need to add attributes
# need to add include callback?
# need to include that these are of class Face
for i in range(lines.shape[0]):
box = Box(path=path)
box.xtl = max(int(lines[i, 1]), 0) * scale
box.ytl = max(int(lines[i, 2]), 0) * scale
box.xbr = max(int(lines[i, 3]), 0) * scale
box.ybr = max(int(lines[i, 4]), 0) * scale
box.occluded = int(0)
box.outside = int(0)
box.generated = int(1)
box.frame = int(lines[i, 0])
logger.debug("Received box {0}".format(str(box.getbox())))
"""
# TODO: May need to include this if attributes load funny
attributes = {}
for label in video.labels:
attributes[label.id] = dict((a.id, a.text) for a in label.attributes)
for attributeid, timeline in attributes.items():
attribute = session.query(Attribute).get(attributeid)
for frame, value in timeline.items():
aa = AttributeAnnotation()
aa.attribute = attribute
aa.frame = frame
aa.value = value
path.attributes.append(aa)
"""
job.paths.append(path)
#print("Appended path %d"%(newID))
#for p in job.paths:
# session.delete(p)
session.add(job)
session.commit()
return min(thresh, num_faces)
def savejob(id, tracks):
job = session.query(Job).get(id)
# Update current job
for path in job.paths:
session.delete(path)
session.commit()
for path in readpaths(tracks):
logger.info(path)
job.paths.append(path)
session.add(job)
session.commit()
# Update neigboring segments
video = job.segment.video
prevseg, nextseg = video.getsegmentneighbors(job.segment)
mergesegments = [
s for s in [prevseg, job.segment, nextseg] if s is not None
]
updatesegments = [s for s in [prevseg, nextseg] if s is not None]
# Create list of merged boxes with given label and userid
labeledboxes = []
for boxes, paths in merge.merge(mergesegments, threshold=0.8):
path = paths[0]
for p in paths:
if p.job.segmentid == job.segmentid:
path = p
break
labeledboxes.append((path.label, path.userid, boxes))
# Remove paths in neighboring segments
for segment in updatesegments:
for path in segment.paths:
session.delete(path)
session.commit()
# Add merged paths to neighboring segments
for label, userid, boxes in labeledboxes:
frames = sorted([box.frame for box in boxes])
for segment in updatesegments:
for job in segment.jobs:
path = Path()
path.label = label
path.userid = userid
addedbox = False
for box in boxes:
if segment.start <= box.frame <= segment.stop:
newbox = Box(path=path)
newbox.frombox(box)
if not box.lost:
addedbox = True
# Some segments and paths might not overlap
if addedbox:
# Add in first frame if it's missing
if (frames[0] < segment.start < frames[-1]
and segment.start not in frames):
newbox = Box(path=path)
newbox.generated = False
newbox.frombox([
box for box in LinearFill(boxes)
if box.frame == segment.start
][0])
job.paths.append(path)
session.add(job)
session.commit()
def __call__(self, args):
if args.load:
for data in pickle.load(open(args.load)):
worker = Worker.lookup(data[0])
worker.numsubmitted = data[1]
worker.numacceptances = data[2]
worker.numrejections = data[3]
worker.blocked = data[4]
worker.donatedamount = data[5]
worker.bonusamount = data[6]
worker.verified = data[7]
print "Loaded {0}".format(worker.id)
session.add(worker)
session.commit()
elif args.dump:
data = []
for worker in session.query(Worker):
data.append((worker.id,
worker.numsubmitted,
worker.numacceptances,
worker.numrejections,
worker.blocked,
worker.donatedamount,
worker.bonusamount,
worker.verified))
print "Dumped {0}".format(worker.id)
pickle.dump(data, open(args.dump, "w"))
elif args.block:
worker = Worker.lookup(args.block)
worker.block("Poor quality work.")
session.add(worker)
session.commit()
print "Blocked {0}".format(args.block)
elif args.unblock:
worker = Worker.lookup(args.unblock)
worker.unblock("Continue working.")
session.add(worker)
session.commit()
print "Unblocked {0}".format(args.unblock)
elif args.search:
query = session.query(Worker)
query = query.filter(Worker.id.like(args.search + "%"))
if query.count():
print "Matches:"
for worker in query:
print worker.id
else:
print "No matches."
elif args.summary:
query = session.query(Worker)
query = session.filter(Worker.id == args.summary)
if query.count():
worker = query.one()
print "Submitted: {0}".format(worker.numsubmitted)
print "Accepted: {0}".format(worker.numacceptances)
print "Rejected: {0}".format(worker.numrejections)
print "Bonuses: {0}".format(worker.bonusamount)
print "Donated: {0}".format(worker.donatedamount)
print "Verified: {0}".format(worker.verified)
print "Blocked: {0}".format(worker.blocked)
else:
print "No matches."
else:
workers = session.query(Worker)
workers = workers.order_by(Worker.numacceptances)
for worker in workers:
extra = ""
if worker.blocked:
extra = "BLOCKED"
if worker.verified:
extra = extra + " VERIFIED"
extra = extra.strip()
data = (worker.id,
worker.numsubmitted,
worker.numacceptances,
worker.numrejections,
extra)
print "{0:<15} {1:>5} jobs {2:>5} acc {3:>5} rej {4}".format(*data)
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."