def track_features(self, timeholder, visitor_data, channel_regions,
position, outdir):
shutil.rmtree(outdir, True)
makedirs(outdir)
for tracks in visitor_data.itervalues():
for startid, event_data in tracks.iteritems():
if startid in ('_full_tracks', '_current_branch'):
continue
for chname, region in channel_regions.iteritems():
for region_name, feature_names in region.iteritems():
try:
frame, obj_label, branch = Tracker.split_nodeid(
startid)
except ValueError:
frame, obj_label = Tracker.split_nodeid(startid)
branch = 1
filename = 'features__P%s__T%05d__O%04d__B%02d__C%s__R%s.txt' \
%(position, frame, obj_label, branch, chname, region_name)
filename = join(outdir, filename)
self._data_per_channel(timeholder, event_data,
filename, chname, region_name,
feature_names, position)
def full_tracks(self, timeholder, visitor_data, position, outdir):
shutil.rmtree(outdir, True)
makedirs(outdir)
for start_id, data in visitor_data.iteritems():
for idx, track in enumerate(data['_full']):
has_header = False
line1 = []
line2 = []
line3 = []
frame, obj_label= Tracker.split_nodeid(start_id)[:2]
filename = 'P%s__T%05d__O%04d__B%02d.txt' \
%(position, frame, obj_label, idx+1)
f = file(join(outdir, filename), 'w')
for node_id in track:
frame, obj_id = Tracker.split_nodeid(node_id)
coordinate = Coordinate(position=position, time=frame)
prefix = [frame, self.meta_data.get_timestamp_relative(coordinate), obj_id]
prefix_names = ['frame', 'time', 'objID']
items = []
for channel in timeholder[frame].values():
for region_id in channel.region_names():
region = channel.get_region(region_id)
if obj_id in region:
flkp = self._map_feature_names(region.feature_names)
if not has_header:
keys = ['classLabel', 'className']
if channel.NAME == 'Primary':
keys += ['centerX', 'centerY']
keys += flkp.keys()
line1 += [channel.NAME.upper()] * len(keys)
line2 += [str(region_id)] * len(keys)
line3 += keys
obj = region[obj_id]
features = region.features_by_name(obj_id, flkp.values())
values = [x if not x is None else '' for x in [obj.iLabel, obj.strClassName]]
if channel.NAME == 'Primary':
values += [obj.oCenterAbs[0], obj.oCenterAbs[1]]
values += list(features)
items.extend(values)
if not has_header:
has_header = True
prefix_str = [''] * len(prefix)
line1 = prefix_str + line1
line2 = prefix_str + line2
line3 = prefix_names + line3
f.write('%s\n' %CSVParams.sep.join(line1))
f.write('%s\n' %CSVParams.sep.join(line2))
f.write('%s\n' %CSVParams.sep.join(line3))
f.write('%s\n' %CSVParams.sep.join([str(i) for i in prefix + items]))
f.close()
def start_nodes(self):
"""Return all start nodes i.e. nodes without incoming edges."""
start_nodes = [
node_id for node_id in self.graph.node_list()
if self.graph.in_degree(node_id) == 0
]
start_nodes.sort(key=lambda x: Tracker.split_nodeid(x)[0])
return start_nodes
def track_features(self, timeholder, visitor_data, channel_regions,
position, outdir):
shutil.rmtree(outdir, True)
makedirs(outdir)
for tracks in visitor_data.itervalues():
for startid, event_data in tracks.iteritems():
if startid.startswith('_'):
continue
for chname, region in channel_regions.iteritems():
for region_name, feature_names in region.iteritems():
try:
frame, obj_label, branch = Tracker.split_nodeid(startid)
except ValueError:
frame, obj_label = Tracker.split_nodeid(startid)
branch = 1
filename = 'features_P%s_T%05d_O%04d_B%02d_C%s_R%s.txt' \
%(position, frame, obj_label, branch, chname, region_name)
filename = join(outdir, filename)
self._data_per_channel(timeholder, event_data, filename, chname,
region_name, feature_names, position)
def centers(self):
"""Return the a list of the object centers for each track."""
centers = dict()
for startid, eventdata in self.iterevents():
if startid in ['_full_tracks', '_current_branch']:
continue
data = list()
for nodeids in zip(*eventdata['tracks']):
for nodeid in nodeids:
obj = self.graph.node_data(nodeid)
frame = Tracker.split_nodeid(nodeid)[0]
data.append((int(frame), obj.iId, obj.oCenterAbs))
centers[startid] = data
return centers
def centers(self):
"""Return the a list of the object centers for each track."""
centers = dict()
for startid, eventdata in self.iterevents():
if startid in ['_full_tracks', '_current_branch']:
continue
data = list()
for nodeids in zip(*eventdata['tracks']):
for nodeid in nodeids:
obj = self.graph.node_data(nodeid)
frame = Tracker.split_nodeid(nodeid)[0]
data.append((int(frame), obj.iId, obj.oCenterAbs))
centers[startid] = data
return centers
def bboxes(self, size=None, border=0):
bboxes = {}
for startid, eventdata in self.iterevents():
if startid in ['_full', '_current']:
continue
data = []
for nodeids in zip(*eventdata['tracks']):
nodeid = nodeids[0]
frame = Tracker.split_nodeid(nodeid)[0]
objids = [Tracker.split_nodeid(n)[1] for n in nodeids]
objects = [self.graph.node_data(n) for n in nodeids]
minX = min([obj.oRoi.upperLeft[0] for obj in objects])
minY = min([obj.oRoi.upperLeft[1] for obj in objects])
maxX = max([obj.oRoi.lowerRight[0] for obj in objects])
maxY = max([obj.oRoi.lowerRight[1] for obj in objects])
width = maxX - minX + 1
height = maxY - minY + 1
centerX = int(
round(np.average([obj.oCenterAbs[0] for obj in objects])))
centerY = int(
round(np.average([obj.oCenterAbs[1] for obj in objects])))
data.append((frame, centerX, centerY, width, height, objids))
data1 = np.array(data, 'O')
if not size is None and len(size) == 2:
diffX = int(size[0] / 2)
diffY = int(size[1] / 2)
else:
diffX = int(max(data1[:, 3]) / 2 + border)
diffY = int(max(data1[:, 4]) / 2 + border)
# convert to float to for numpy float64 type
timedata = [(int(d[0]), (d[1] - diffX, d[2] - diffY,
d[1] + diffX - 1 + size[0] % 2,
d[2] + diffY - 1 + size[1] % 2), d[5])
for d in data1]
bboxes[startid] = timedata
return bboxes
def bboxes(self, size=None, border=0):
bboxes = {}
for startid, eventdata in self.iterevents():
if startid in ['_full', '_current']:
continue
data = []
for nodeids in zip(*eventdata['tracks']):
nodeid = nodeids[0]
frame = Tracker.split_nodeid(nodeid)[0]
objids = [Tracker.split_nodeid(n)[1] for n in nodeids]
objects = [self.graph.node_data(n) for n in nodeids]
minX = min([obj.oRoi.upperLeft[0] for obj in objects])
minY = min([obj.oRoi.upperLeft[1] for obj in objects])
maxX = max([obj.oRoi.lowerRight[0] for obj in objects])
maxY = max([obj.oRoi.lowerRight[1] for obj in objects])
width = maxX - minX + 1
height = maxY - minY + 1
centerX = int(round(np.average([obj.oCenterAbs[0] for obj in objects])))
centerY = int(round(np.average([obj.oCenterAbs[1] for obj in objects])))
data.append((frame, centerX, centerY, width, height, objids))
data1 = np.array(data, 'O')
if not size is None and len(size) == 2:
diffX = int(size[0] / 2)
diffY = int(size[1] / 2)
else:
diffX = int(max(data1[:,3])/2 + border)
diffY = int(max(data1[:,4])/2 + border)
# convert to float to for numpy float64 type
timedata = [(int(d[0]),
(d[1] - diffX,
d[2] - diffY,
d[1] + diffX - 1 + size[0] %2,
d[2] + diffY - 1 + size[1] %2),
d[5]) for d in data1]
bboxes[startid] = timedata
return bboxes
def _split_nodeid(self, nodeid):
ret = Tracker.split_nodeid(nodeid)
if len(ret) == 2:
ret = ret + (1,)
return ret
def _data_per_channel(self, timeholder, event_data, filename, channel_name, region_name,
feature_names, position):
eventid = event_data['eventId']
event_frame, _ = Tracker.split_nodeid(eventid)
has_split = 'splitId' in event_data
header_names = ['Frame', 'Timestamp', 'isEvent']
if has_split:
header_names.append('isSplit')
if event_data['splitId'] is not None:
split_frame, _ = Tracker.split_nodeid(event_data['splitId'])
else:
split_frame = None
table = []
# zip nodes with same time together
for nodeids in zip(*event_data['tracks']):
objids = []
frame = None
for nodeid in nodeids:
node_frame, objid = Tracker.split_nodeid(nodeid)
if frame is None:
frame = node_frame
else:
assert frame == node_frame
objids.append(objid)
channel = timeholder[frame][channel_name]
sample_holder = channel.get_region(region_name)
if feature_names is None:
feature_names = sample_holder.feature_names
if CSVParams.objId not in header_names:
# setup header line
header_names.append(CSVParams.objId)
header_names += [CSVParams.class_ %x for x in ['name', 'label', 'probability']]
# only feature_names scales according to settings
header_names += [CSVParams.feature %fn for fn in feature_names]
header_names += [CSVParams.tracking %tf for tf in CSVParams.tracking_features]
coordinate = Coordinate(position=position, time=frame)
data = {'Frame' : frame,
'Timestamp': self.meta_data.get_timestamp_relative(coordinate),
'isEvent': int(frame==event_frame)}
if has_split:
data['isSplit'] = int(frame==split_frame)
#for iIdx, iObjId in enumerate(lstObjectIds):
objid = objids[0]
if objid in sample_holder:
sample = sample_holder[objid]
data[CSVParams.objId] = objid
# classification data
if sample.iLabel is not None:
data[CSVParams.class_ %'label'] = sample.iLabel
data[CSVParams.class_ %'name'] = sample.strClassName
data[CSVParams.class_ %'probability'] = \
','.join(['%d:%.5f' % (int(x),y) for x,y in sample.dctProb.iteritems()])
common_ftr = [f for f in set(sample_holder.feature_names).intersection(feature_names)]
features = sample_holder.features_by_name(objid, common_ftr)
for feature, fname in zip(features, common_ftr):
data[CSVParams.feature %fname] = feature
# features not calculated are exported as NAN
diff_ftr = [f for f in set(feature_names).difference(sample_holder.feature_names)]
for df in diff_ftr:
data[CSVParams.feature %df] = float("NAN")
# object tracking data (absolute center)
data[CSVParams.tracking %'center_x'] = sample.oCenterAbs[0]
data[CSVParams.tracking %'center_y'] = sample.oCenterAbs[1]
data[CSVParams.tracking %'upperleft_x'] = sample.oRoi.upperLeft[0]
data[CSVParams.tracking %'upperleft_y'] = sample.oRoi.upperLeft[1]
data[CSVParams.tracking %'lowerright_x'] = sample.oRoi.lowerRight[0]
data[CSVParams.tracking %'lowerright_y'] = sample.oRoi.lowerRight[1]
else:
# we rather skip the entire event in case the object ID is not valid
return
table.append(data)
if len(table) > 0:
with open(filename, 'w') as fp:
writer = csv.DictWriter(fp, fieldnames=header_names,
delimiter=CSVParams.sep)
writer.writeheader()
writer.writerows(table)
def start_nodes(self):
"""Return all start nodes i.e. nodes without incoming edges."""
start_nodes = [node_id for node_id in self.graph.node_list()
if self.graph.in_degree(node_id) == 0]
start_nodes.sort(key=lambda x: Tracker.split_nodeid(x)[0])
return start_nodes
def __init__(self, eventselector, strPathIn, oP, strPathOut,
imageCompression="85",
imageSuffix=".jpg",
border=0,
writeSubdirs=True,
writeDescription=True,
size=None,
oneFilePerTrack=False):
self._bHasImages = False
dctTimePoints = {}
for strStartId, lstTimeData in eventselector.bboxes( \
size=size, border=border).iteritems():
items = Tracker.split_nodeid(strStartId)
iStartT, iObjId = items[:2]
if len(items) == 3:
branch_id = items[2]
else:
branch_id = 1
if writeSubdirs:
strPathOutEvent = os.path.join(strPathOut,
self._format_name(oP, iStartT, iObjId, branch_id))
else:
strPathOutEvent = strPathOut
makedirs(strPathOutEvent)
if writeDescription:
oFile = file(os.path.join(strPathOutEvent,
"_%s.txt" % self._format_name(oP, iStartT, iObjId, branch_id)), "w")
lstData = ["Frame", "ObjId", "x1", "y1", "x2", "y2"]
oFile.write("%s\n" % "\t".join(map(str, lstData)))
for iCnt, (iT, tplBoundingBox, lstObjIds) in enumerate(lstTimeData):
if writeDescription:
lstData = [iT, ';'.join(map(str, lstObjIds))] + list(tplBoundingBox)
oFile.write("%s\n" % "\t".join(map(str, lstData)))
if not iT in dctTimePoints:
dctTimePoints[iT] = []
dctTimePoints[iT].append((strStartId, lstObjIds, iCnt, strPathOutEvent, tplBoundingBox))
if writeDescription:
oFile.close()
for idx, (iT, lstItems) in enumerate(dctTimePoints.iteritems()):
#print iT, lstItems
imgXY = self._getImage(strPathIn, iT)
for strStartId, lstObjIds, iCnt, strPathOutEvent, tplBoundingBox in lstItems:
x1, y1, x2, y2 = tplBoundingBox
x1Corr = 0 if x1 < 0 else x1
y1Corr = 0 if y1 < 0 else y1
x2Corr = imgXY.width-1 if x2 >= imgXY.width else x2
y2Corr = imgXY.height-1 if y2 >= imgXY.height else y2
imgSub = ccore.subImage(imgXY,
ccore.Diff2D(x1Corr, y1Corr),
ccore.Diff2D(x2Corr-x1Corr+1, y2Corr-y1Corr+1))
if (x1 < 0 or y1 < 0 or
x2 >= imgXY.width or y2 >= imgXY.height):
imgSub2 = self.IMAGE_CLASS(size[0], size[1])
ccore.copySubImage(imgSub, imgSub2, ccore.Diff2D(x1Corr-x1, y1Corr-y1))
imgSub = imgSub2
assert imgSub.width == size[0]
assert imgSub.width == x2-x1+1
assert imgSub.height == size[1]
assert imgSub.height == y2-y1+1
if self.PROCESS_LABEL:
lstImages = []
for iObjId in lstObjIds:
lstImages.append(ccore.copyImageIfLabel(imgSub, imgSub, iObjId))
imgSub = ccore.projectImage(lstImages, ccore.ProjectionType.MaxProjection)
strFilenameImage = os.path.join(strPathOutEvent, "P%s__T%05d%s" % (oP, iT, imageSuffix))
ccore.writeImage(imgSub, strFilenameImage)
if oneFilePerTrack and os.path.isdir(strPathOut):
self.convertToOneFilePerTrack(strPathOut, imageCompression)
def full_tracks(self, timeholder, visitor_data, position, outdir):
shutil.rmtree(outdir, True)
makedirs(outdir)
for start_id, data in visitor_data.iteritems():
for idx, track in enumerate(data['_full_tracks']):
has_header = False
line1 = []
line2 = []
line3 = []
frame, obj_label = Tracker.split_nodeid(start_id)[:2]
filename = 'P%s__T%05d__O%04d__B%02d.txt' \
%(position, frame, obj_label, idx+1)
f = file(join(outdir, filename), 'w')
for node_id in track:
frame, obj_id = Tracker.split_nodeid(node_id)
coordinate = Coordinate(position=position, time=frame)
prefix = [
frame,
self.meta_data.get_timestamp_relative(coordinate),
obj_id
]
prefix_names = ['frame', 'time', 'objID']
items = []
for channel in timeholder[frame].values():
for region_id in channel.region_names():
region = channel.get_region(region_id)
if obj_id in region:
flkp = self._map_feature_names(
region.feature_names)
if not has_header:
keys = ['classLabel', 'className']
if channel.NAME == 'Primary':
keys += ['centerX', 'centerY']
keys += flkp.keys()
line1 += [channel.NAME.upper()] * len(keys)
line2 += [str(region_id)] * len(keys)
line3 += keys
obj = region[obj_id]
features = region.features_by_name(
obj_id, flkp.values())
values = [
x if not x is None else ''
for x in [obj.iLabel, obj.strClassName]
]
if channel.NAME == 'Primary':
values += [
obj.oCenterAbs[0], obj.oCenterAbs[1]
]
values += list(features)
items.extend(values)
if not has_header:
has_header = True
prefix_str = [''] * len(prefix)
line1 = prefix_str + line1
line2 = prefix_str + line2
line3 = prefix_names + line3
f.write('%s\n' % CSVParams.sep.join(line1))
f.write('%s\n' % CSVParams.sep.join(line2))
f.write('%s\n' % CSVParams.sep.join(line3))
f.write(
'%s\n' %
CSVParams.sep.join([str(i) for i in prefix + items]))
f.close()
def _data_per_channel(self, timeholder, event_data, filename, channel_name,
region_name, feature_names, position):
eventid = event_data['eventId']
event_frame, _ = Tracker.split_nodeid(eventid)
has_split = 'splitId' in event_data
header_names = ['Frame', 'Timestamp', 'isEvent']
if has_split:
header_names.append('isSplit')
if event_data['splitId'] is not None:
split_frame, _ = Tracker.split_nodeid(event_data['splitId'])
else:
split_frame = None
table = []
# zip nodes with same time together
for nodeids in zip(*event_data['tracks']):
objids = []
frame = None
for nodeid in nodeids:
node_frame, objid = Tracker.split_nodeid(nodeid)
if frame is None:
frame = node_frame
else:
assert frame == node_frame
objids.append(objid)
channel = timeholder[frame][channel_name]
sample_holder = channel.get_region(region_name)
if feature_names is None:
feature_names = sample_holder.feature_names
if CSVParams.objId not in header_names:
# setup header line
header_names.append(CSVParams.objId)
header_names += [
CSVParams.class_ % x
for x in ['name', 'label', 'probability']
]
# only feature_names scales according to settings
header_names += [
CSVParams.feature % fn for fn in feature_names
]
header_names += [
CSVParams.tracking % tf
for tf in CSVParams.tracking_features
]
coordinate = Coordinate(position=position, time=frame)
data = {
'Frame': frame,
'Timestamp': self.meta_data.get_timestamp_relative(coordinate),
'isEvent': int(frame == event_frame)
}
if has_split:
data['isSplit'] = int(frame == split_frame)
#for iIdx, iObjId in enumerate(lstObjectIds):
objid = objids[0]
if objid in sample_holder:
sample = sample_holder[objid]
data[CSVParams.objId] = objid
# classification data
if sample.iLabel is not None:
data[CSVParams.class_ % 'label'] = sample.iLabel
data[CSVParams.class_ % 'name'] = sample.strClassName
data[CSVParams.class_ %'probability'] = \
','.join(['%d:%.5f' % (int(x),y) for x,y in
sample.dctProb.iteritems()])
common_ftr = [
f for f in set(sample_holder.feature_names).intersection(
feature_names)
]
features = sample_holder.features_by_name(objid, common_ftr)
for feature, fname in zip(features, common_ftr):
data[CSVParams.feature % fname] = feature
# features not calculated are exported as NAN
diff_ftr = [
f for f in set(feature_names).difference(
sample_holder.feature_names)
]
for df in diff_ftr:
data[CSVParams.feature % df] = float("NAN")
# object tracking data (absolute center)
data[CSVParams.tracking % 'center_x'] = sample.oCenterAbs[0]
data[CSVParams.tracking % 'center_y'] = sample.oCenterAbs[1]
data[CSVParams.tracking %
'upperleft_x'] = sample.oRoi.upperLeft[0]
data[CSVParams.tracking %
'upperleft_y'] = sample.oRoi.upperLeft[1]
data[CSVParams.tracking %
'lowerright_x'] = sample.oRoi.lowerRight[0]
data[CSVParams.tracking %
'lowerright_y'] = sample.oRoi.lowerRight[1]
else:
# we rather skip the entire event in case the object ID is not valid
return
table.append(data)
if len(table) > 0:
with open(filename, 'wb') as fp:
writer = csv.DictWriter(fp,
fieldnames=header_names,
delimiter=CSVParams.sep)
writer.writeheader()
writer.writerows(table)
def _split_nodeid(self, nodeid):
ret = Tracker.split_nodeid(nodeid)
if len(ret) == 2:
ret = ret + (1, )
return ret
