def main(doTubeFormat, inFileName, outTrajIDfileName, featPath, nrTrajThresh4Tube=0):
tube2trajIDs.check()
outPath = os.path.dirname(outTrajIDfileName)
if not os.path.exists( outPath ):
os.makedirs(outPath)
# read trajectory positions
print '\tGet trajectory positions;',
geoFeat = denseTraj.getFeatFromFileByName(featPath, 'geo')
# if proposals are stored as Tubes already, read them
if doTubeFormat:
tubeProposals = TubeList()
tubeProposals.readHDF5(inFileName)
# initialize with a single name
inHDF5fileClusts = ['tubes']
else:
print '\tGet video dimensions;',
vidInfo = denseTraj.getFeatFromFileByName(featPath, 'vidinfo')
xmax = vidInfo[1]
ymax = vidInfo[2]
# input cluster file
inHDF5fileClusts = h5py.File( inFileName, 'r')
# go over all features
totNrProposals = 0
for featName in inHDF5fileClusts:
if not doTubeFormat:
# get the clustered proposals for this feature type in SLINK pointer-representation
dset = inHDF5fileClusts[featName]
mergedTracks = dset[()]
# convert pointer-representation to tube proposals
if len(mergedTracks.shape) > 1 and mergedTracks.shape[0] > 1:
tubeProposals = denseTraj.createProposals(mergedTracks, geoFeat, nrTrajThresh4Tube, xmax, ymax)
nrProposals = len(tubeProposals)
print '\tFeature: "%s", number of proposals: %d;' % (featName,nrProposals)
# if ok
if nrProposals > 0:
# write as traj IDs
if totNrProposals==0:
outHDF5file = h5py.File( outTrajIDfileName, 'w', compression="gzip", compression_opts=9)
for j in range(nrProposals):
trajIDs = tubeProposals[j].tube2trajIDs(geoFeat)
outHDF5file.create_dataset(str(totNrProposals), data=trajIDs)
totNrProposals += 1
print '\tWrite trajectory IDs to', outTrajIDfileName
outHDF5file.close()
def main(inFileName, featPath, nrTrajThresh4Tube, outFileName):
#print inFileName, featPath, nrTrajThresh4Tube, outFileName
#sys.exit()
outPath = os.path.dirname(outFileName)
if not os.path.exists(outPath):
os.makedirs(outPath)
# output file
print '\tWrite tubes to: %s' % (outFileName)
outHDF5fileTubes = h5py.File(outFileName, 'w')
# get the trajectory locations
geoFeat = denseTraj.getFeatFromFileByName(featPath, 'geo')
# get vid info (length, width, height)
vidInfo = denseTraj.getFeatFromFileByName(featPath, 'vidinfo')
xmax = vidInfo[1]
ymax = vidInfo[2]
# input cluster file
inHDF5fileClusts = h5py.File(inFileName, 'r')
propOutstartID = 0
for featName in inHDF5fileClusts:
# get the clustered proposals for this feature type in SLINK pointer-representation
#featName = featNames[featNameID]
#fnameSS = os.path.join(inPath, featName + suffixIn + '.npy')
#mergedTracks = np.load( fnameSS )
dset = inHDF5fileClusts[featName]
mergedTracks = dset[()]
# convert pointer-representation to tube proposals and write to disk
if len(mergedTracks.shape) > 1 and mergedTracks.shape[0] > 1:
tubeProposals = denseTraj.createProposals(mergedTracks, geoFeat,
nrTrajThresh4Tube, xmax,
ymax)
nrProposals = len(tubeProposals)
print '\tFeature: "%s"; Number of proposals: %d' % (featName,
nrProposals),
# if OK
if nrProposals > 0:
# write to file
#print '; writing proposals:', outPath
tubeProposals.writeHDF5(outHDF5fileTubes, propOutstartID)
propOutstartID += nrProposals
print 'tot:', propOutstartID
# close the file
outHDF5fileTubes.close()
def main(inFileName, featPath, nrTrajThresh4Tube, outFileName):
# print inFileName, featPath, nrTrajThresh4Tube, outFileName
# sys.exit()
outPath = os.path.dirname(outFileName)
if not os.path.exists(outPath):
os.makedirs(outPath)
# output file
print "\tWrite tubes to: %s" % (outFileName)
outHDF5fileTubes = h5py.File(outFileName, "w")
# get the trajectory locations
geoFeat = denseTraj.getFeatFromFileByName(featPath, "geo")
# get vid info (length, width, height)
vidInfo = denseTraj.getFeatFromFileByName(featPath, "vidinfo")
xmax = vidInfo[1]
ymax = vidInfo[2]
# input cluster file
inHDF5fileClusts = h5py.File(inFileName, "r")
propOutstartID = 0
for featName in inHDF5fileClusts:
# get the clustered proposals for this feature type in SLINK pointer-representation
# featName = featNames[featNameID]
# fnameSS = os.path.join(inPath, featName + suffixIn + '.npy')
# mergedTracks = np.load( fnameSS )
dset = inHDF5fileClusts[featName]
mergedTracks = dset[()]
# convert pointer-representation to tube proposals and write to disk
if len(mergedTracks.shape) > 1 and mergedTracks.shape[0] > 1:
tubeProposals = denseTraj.createProposals(mergedTracks, geoFeat, nrTrajThresh4Tube, xmax, ymax)
nrProposals = len(tubeProposals)
print '\tFeature: "%s"; Number of proposals: %d' % (featName, nrProposals),
# if OK
if nrProposals > 0:
# write to file
# print '; writing proposals:', outPath
tubeProposals.writeHDF5(outHDF5fileTubes, propOutstartID)
propOutstartID += nrProposals
print "tot:", propOutstartID
# close the file
outHDF5fileTubes.close()
def main(doTubeFormat, inFileName, outIoUfile, gtPath, nrTrajThresh4Tube=-1, featPath=''):
# check if cython is used
tubeIoU.check()
# read ground truth tubes
print '\tRead ground truth;',
gtTubes = TubeList()
gtTubes.readHDF5(gtPath)
outPath = os.path.dirname(outIoUfile)
if not os.path.exists( outPath ):
os.makedirs(outPath)
# if proposals are stored as Tubes already, read them
if doTubeFormat:
tubeProposals = TubeList()
tubeProposals.readHDF5(inFileName)
# initialize with a single name
inHDF5fileClusts = ['tubes']
else:
# otherwise, read trajectory positions and vidinfo to generate proposals later
print '\tGet trajectory positions;',
geoFeat = denseTraj.getFeatFromFileByName(featPath, 'geo')
print '\tGet video dimensions;',
vidInfo = denseTraj.getFeatFromFileByName(featPath, 'vidinfo')
xmax = vidInfo[1]
ymax = vidInfo[2]
# input cluster file
inHDF5fileClusts = h5py.File( inFileName, 'r')
aboMax = []
# go over all features
for featName in inHDF5fileClusts:
if not doTubeFormat:
# get the clustered proposals for this feature type in SLINK pointer-representation
dset = inHDF5fileClusts[featName]
mergedTracks = dset[()]
# convert pointer-representation to tube proposals
if len(mergedTracks.shape) > 1 and mergedTracks.shape[0] > 1:
tubeProposals = denseTraj.createProposals(mergedTracks, geoFeat, nrTrajThresh4Tube, xmax, ymax)
nrProposals = len(tubeProposals)
print '\tFeature: "%s", number of proposals: %d;' % (featName,nrProposals),
# if ok
if nrProposals > 0:
# get Intersection over Union scores
aboMatCur = tubeProposals.computeTubeOverlap(gtTubes)
# if first feature, write to new file otherwise append to existing file
if aboMax == []:
fileOut = open(outIoUfile, 'w')
else:
fileOut = open(outIoUfile, 'a')
# write scores
for prop in range(aboMatCur.shape[0]):
for score in aboMatCur[prop,:]:
fileOut.write('%f ' % score)
fileOut.write('\n')
fileOut.close()
# keep track of the maximum score
curMax = np.max(aboMatCur, axis=0)
if aboMax == []:
aboMax = curMax
else:
aboMax = np.maximum( aboMax, curMax)
print 'Best IoUs:', curMax, '; Best so far:', aboMax
print '\tWriting IoU scores to', outIoUfile
if not doTubeFormat:
inHDF5fileClusts.close()
def main(featPath, nrSpatNeighbors, isTrimmedVideo, outFileName):
"""main function"""
if isTrimmedVideo:
featNorm = "Z"
tempScale = 0.0
else:
featNorm = "MiMaN"
tempScale = 1.0
# simple timing
tic = time.time()
# a bool to check if the spatial neighborhood is already computed
spatFeatComputed = False
# a bool to check if there were enough trajectories to cluster
enoughTrajectories = True
# mapping from feature to similarities (to avoid recomputing)
feat2neighborSims = {}
# if some features are already computed, only compute missing features
doneFeatNames = []
if not os.path.exists(outFileName):
outHDF5file = h5py.File( outFileName, 'w')
outHDF5file.close()
print '\twriting features to %s' % outFileName
else:
try:
print '\tadding (possible) missing features to %s' % outFileName
inHDF5file = h5py.File( outFileName, 'r')
doneFeatNames = inHDF5file.keys()
inHDF5file.close()
except IOError as e:
outHDF5file = h5py.File( outFileName, 'w')
outHDF5file.close()
print '\toverwriting features to %s' % outFileName
# for a given feature (or feature combination, a combination has a '-' in it)
for featName in featNames:
#print 'featName "%s"' % featName
# if featName already computed, skip it.
if featName not in doneFeatNames:
if enoughTrajectories:
#print 'do', featName
#sys.exit()
#print featName.split('-')
if not spatFeatComputed:
feat = denseTraj.getFeatFromFileByName(featPath, 'spat', tempScale)
nrTracks = feat.shape[0]
if nrTracks < nrSpatNeighbors:
print featName, 'not enough tracks to continue:', nrTracks,
enoughTrajectories = False
continue
#vidInfo = denseTraj.getFeatFromFileByName(featPath, 'vidinfo')
#simMatIDs, simMat = tubeClust.getTrackNeighborsYaelYnumpyChunked(feat, nrSpatNeighbors, int(vidInfo[0]))
simMatIDs, _simMat = tubeClust.getTrackNeighborsYaelYnumpy(feat, nrSpatNeighbors)
neighborGraph = tubeClust.getNeighborGraph(simMatIDs, nrTracks)
spatFeatComputed = True
if enoughTrajectories:
# combined features are separated with a dash "-"
combiFeatNames = featName.split('-')
# first do the first feature
if combiFeatNames[0] not in feat2neighborSims.keys():
feat = denseTraj.getFeatFromFileByName(featPath, combiFeatNames[0], tempScale)
neighborSims = tubeClust.getNeighborSimilarities(neighborGraph, feat, featName2DistFun[combiFeatNames[0]])
if featNorm == 'Z':
neighborSims = tubeClust.normalizeNeighborSimilarities(neighborSims)
if featNorm == 'MiMaN':
neighborSims = tubeClust.normalizeNeighborSimilaritiesMiMa(neighborSims)
feat2neighborSims[ combiFeatNames[0] ] = neighborSims
else:
neighborSims = feat2neighborSims[ combiFeatNames[0] ]
# add the scores of the to-be-combined features
for indiFeat in combiFeatNames[1:]:
#print indiFeat,
if indiFeat not in feat2neighborSims.keys():
feat = denseTraj.getFeatFromFileByName(featPath, indiFeat, tempScale)
neighborSimsIndi = tubeClust.getNeighborSimilarities(neighborGraph, feat, featName2DistFun[indiFeat] )
if featNorm == 'Z':
neighborSimsIndi = tubeClust.normalizeNeighborSimilarities(neighborSimsIndi)
if featNorm == 'MiMaN':
neighborSimsIndi = tubeClust.normalizeNeighborSimilaritiesMiMa(neighborSimsIndi)
feat2neighborSims[indiFeat] = neighborSimsIndi
else:
neighborSimsIndi = feat2neighborSims[indiFeat]
neighborSims = neighborSims * neighborSimsIndi
Lambda,Pi = slink.SSclustSlink(neighborGraph, neighborSims, nrTracks)
mergedTracks = tubeClust.convertPointerRepresentation(np.array(Lambda), Pi, nrTracks)
toc = time.time()
print '; Time: %.2fs' % (toc-tic)
if not enoughTrajectories:
mergedTracks = -1
print '\twrite to: %s, "%s"' % (outFileName, featName)
outHDF5file = h5py.File( outFileName, 'a')
outHDF5file.create_dataset(featName, data=mergedTracks)
outHDF5file.close()
def main(doTubeFormat,
inFileName,
outIoUfile,
gtPath,
nrTrajThresh4Tube=-1,
featPath=''):
# check if cython is used
tubeIoU.check()
# read ground truth tubes
print '\tRead ground truth;',
gtTubes = TubeList()
gtTubes.readHDF5(gtPath)
outPath = os.path.dirname(outIoUfile)
if not os.path.exists(outPath):
os.makedirs(outPath)
# if proposals are stored as Tubes already, read them
if doTubeFormat:
tubeProposals = TubeList()
tubeProposals.readHDF5(inFileName)
# initialize with a single name
inHDF5fileClusts = ['tubes']
else:
# otherwise, read trajectory positions and vidinfo to generate proposals later
print '\tGet trajectory positions;',
geoFeat = denseTraj.getFeatFromFileByName(featPath, 'geo')
print '\tGet video dimensions;',
vidInfo = denseTraj.getFeatFromFileByName(featPath, 'vidinfo')
xmax = vidInfo[1]
ymax = vidInfo[2]
# input cluster file
inHDF5fileClusts = h5py.File(inFileName, 'r')
aboMax = []
# go over all features
for featName in inHDF5fileClusts:
if not doTubeFormat:
# get the clustered proposals for this feature type in SLINK pointer-representation
dset = inHDF5fileClusts[featName]
mergedTracks = dset[()]
# convert pointer-representation to tube proposals
if len(mergedTracks.shape) > 1 and mergedTracks.shape[0] > 1:
tubeProposals = denseTraj.createProposals(
mergedTracks, geoFeat, nrTrajThresh4Tube, xmax, ymax)
nrProposals = len(tubeProposals)
print '\tFeature: "%s", number of proposals: %d;' % (featName,
nrProposals),
# if ok
if nrProposals > 0:
# get Intersection over Union scores
aboMatCur = tubeProposals.computeTubeOverlap(gtTubes)
# if first feature, write to new file otherwise append to existing file
if aboMax == []:
fileOut = open(outIoUfile, 'w')
else:
fileOut = open(outIoUfile, 'a')
# write scores
for prop in range(aboMatCur.shape[0]):
for score in aboMatCur[prop, :]:
fileOut.write('%f ' % score)
fileOut.write('\n')
fileOut.close()
# keep track of the maximum score
curMax = np.max(aboMatCur, axis=0)
if aboMax == []:
aboMax = curMax
else:
aboMax = np.maximum(aboMax, curMax)
print 'Best IoUs:', curMax, '; Best so far:', aboMax
print '\tWriting IoU scores to', outIoUfile
if not doTubeFormat:
inHDF5fileClusts.close()
def main(doTubeFormat,
inFileName,
outTrajIDfileName,
featPath,
nrTrajThresh4Tube=0):
tube2trajIDs.check()
outPath = os.path.dirname(outTrajIDfileName)
if not os.path.exists(outPath):
os.makedirs(outPath)
# read trajectory positions
print '\tGet trajectory positions;',
geoFeat = denseTraj.getFeatFromFileByName(featPath, 'geo')
# if proposals are stored as Tubes already, read them
if doTubeFormat:
tubeProposals = TubeList()
tubeProposals.readHDF5(inFileName)
# initialize with a single name
inHDF5fileClusts = ['tubes']
else:
print '\tGet video dimensions;',
vidInfo = denseTraj.getFeatFromFileByName(featPath, 'vidinfo')
xmax = vidInfo[1]
ymax = vidInfo[2]
# input cluster file
inHDF5fileClusts = h5py.File(inFileName, 'r')
# go over all features
totNrProposals = 0
for featName in inHDF5fileClusts:
if not doTubeFormat:
# get the clustered proposals for this feature type in SLINK pointer-representation
dset = inHDF5fileClusts[featName]
mergedTracks = dset[()]
# convert pointer-representation to tube proposals
if len(mergedTracks.shape) > 1 and mergedTracks.shape[0] > 1:
tubeProposals = denseTraj.createProposals(
mergedTracks, geoFeat, nrTrajThresh4Tube, xmax, ymax)
nrProposals = len(tubeProposals)
print '\tFeature: "%s", number of proposals: %d;' % (featName,
nrProposals)
# if ok
if nrProposals > 0:
# write as traj IDs
if totNrProposals == 0:
outHDF5file = h5py.File(outTrajIDfileName,
'w',
compression="gzip",
compression_opts=9)
for j in range(nrProposals):
trajIDs = tubeProposals[j].tube2trajIDs(geoFeat)
outHDF5file.create_dataset(str(totNrProposals), data=trajIDs)
totNrProposals += 1
print '\tWrite trajectory IDs to', outTrajIDfileName
outHDF5file.close()
def main(featPath, nrSpatNeighbors, isTrimmedVideo, outFileName):
"""main function"""
if isTrimmedVideo:
featNorm = "Z"
tempScale = 0.0
else:
featNorm = "MiMaN"
tempScale = 1.0
# simple timing
tic = time.time()
# a bool to check if the spatial neighborhood is already computed
spatFeatComputed = False
# a bool to check if there were enough trajectories to cluster
enoughTrajectories = True
# mapping from feature to similarities (to avoid recomputing)
feat2neighborSims = {}
# if some features are already computed, only compute missing features
doneFeatNames = []
if not os.path.exists(outFileName):
outHDF5file = h5py.File(outFileName, 'w')
outHDF5file.close()
print '\twriting features to %s' % outFileName
else:
try:
print '\tadding (possible) missing features to %s' % outFileName
inHDF5file = h5py.File(outFileName, 'r')
doneFeatNames = inHDF5file.keys()
inHDF5file.close()
except IOError as e:
outHDF5file = h5py.File(outFileName, 'w')
outHDF5file.close()
print '\toverwriting features to %s' % outFileName
# for a given feature (or feature combination, a combination has a '-' in it)
for featName in featNames:
#print 'featName "%s"' % featName
# if featName already computed, skip it.
if featName not in doneFeatNames:
if enoughTrajectories:
#print 'do', featName
#sys.exit()
#print featName.split('-')
if not spatFeatComputed:
feat = denseTraj.getFeatFromFileByName(
featPath, 'spat', tempScale)
nrTracks = feat.shape[0]
if nrTracks < nrSpatNeighbors:
print featName, 'not enough tracks to continue:', nrTracks,
enoughTrajectories = False
continue
#vidInfo = denseTraj.getFeatFromFileByName(featPath, 'vidinfo')
#simMatIDs, simMat = tubeClust.getTrackNeighborsYaelYnumpyChunked(feat, nrSpatNeighbors, int(vidInfo[0]))
simMatIDs, _simMat = tubeClust.getTrackNeighborsYaelYnumpy(
feat, nrSpatNeighbors)
neighborGraph = tubeClust.getNeighborGraph(
simMatIDs, nrTracks)
spatFeatComputed = True
if enoughTrajectories:
# combined features are separated with a dash "-"
combiFeatNames = featName.split('-')
# first do the first feature
if combiFeatNames[0] not in feat2neighborSims.keys():
feat = denseTraj.getFeatFromFileByName(
featPath, combiFeatNames[0], tempScale)
neighborSims = tubeClust.getNeighborSimilarities(
neighborGraph, feat,
featName2DistFun[combiFeatNames[0]])
if featNorm == 'Z':
neighborSims = tubeClust.normalizeNeighborSimilarities(
neighborSims)
if featNorm == 'MiMaN':
neighborSims = tubeClust.normalizeNeighborSimilaritiesMiMa(
neighborSims)
feat2neighborSims[combiFeatNames[0]] = neighborSims
else:
neighborSims = feat2neighborSims[combiFeatNames[0]]
# add the scores of the to-be-combined features
for indiFeat in combiFeatNames[1:]:
#print indiFeat,
if indiFeat not in feat2neighborSims.keys():
feat = denseTraj.getFeatFromFileByName(
featPath, indiFeat, tempScale)
neighborSimsIndi = tubeClust.getNeighborSimilarities(
neighborGraph, feat, featName2DistFun[indiFeat])
if featNorm == 'Z':
neighborSimsIndi = tubeClust.normalizeNeighborSimilarities(
neighborSimsIndi)
if featNorm == 'MiMaN':
neighborSimsIndi = tubeClust.normalizeNeighborSimilaritiesMiMa(
neighborSimsIndi)
feat2neighborSims[indiFeat] = neighborSimsIndi
else:
neighborSimsIndi = feat2neighborSims[indiFeat]
neighborSims = neighborSims * neighborSimsIndi
Lambda, Pi = slink.SSclustSlink(neighborGraph, neighborSims,
nrTracks)
mergedTracks = tubeClust.convertPointerRepresentation(
np.array(Lambda), Pi, nrTracks)
toc = time.time()
print '; Time: %.2fs' % (toc - tic)
if not enoughTrajectories:
mergedTracks = -1
print '\twrite to: %s, "%s"' % (outFileName, featName)
outHDF5file = h5py.File(outFileName, 'a')
outHDF5file.create_dataset(featName, data=mergedTracks)
outHDF5file.close()
