def darwin(fullfeatnames, st, num_videos, darwins_path):
if not exists(darwins_path):
makedirs(darwins_path)
for feat_t in fullfeatnames:
# node_darwins[feat_t] = dict()
if not exists(join(darwins_path, feat_t)):
makedirs(join(darwins_path, feat_t))
for featname in fullfeatnames[feat_t]:
output_filepath = join(darwins_path, feat_t, basename(featname))
if isfile(output_filepath):
print('%s -> OK' % (featname))
continue
start_time = time.time()
with open(featname, 'rb') as f:
data = cPickle.load(f)
# compute VD
node_darwins = dict()
node_darwins[1] = darwin(data['X'])
for id, X in data['tree_perframe'].iteritems():
node_darwins[id] = darwin(X)
# construct a list of edge pairs for easy access
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(node_darwins=node_darwins), f)
elapsed_time = time.time() - start_time
print('%s -> DONE (in %.2f secs)' %
(output_filepath, elapsed_time))
return None
# ==============================================================================
# Helper functions
# ==============================================================================
def darwin(fullfeatnames, st, num_videos, darwins_path):
if not exists(darwins_path):
makedirs(darwins_path)
for feat_t in fullfeatnames:
# node_darwins[feat_t] = dict()
if not exists(join(darwins_path, feat_t)):
makedirs(join(darwins_path, feat_t))
for featname in fullfeatnames[feat_t]:
output_filepath = join(darwins_path, feat_t, basename(featname))
if isfile(output_filepath):
print('%s -> OK' % (featname))
continue
start_time = time.time()
with open(featname, 'rb') as f:
data = cPickle.load(f)
# compute VD
node_darwins = dict()
node_darwins[1] = darwin(data['X'])
for id, X in data['tree_perframe'].iteritems():
node_darwins[id] = darwin(X)
# construct a list of edge pairs for easy access
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(node_darwins=node_darwins), f)
elapsed_time = time.time() - start_time
print('%s -> DONE (in %.2f secs)' % (output_filepath, elapsed_time))
return None
# ==============================================================================
# Helper functions
# ==============================================================================
def _compute_vd_descriptors(tracklets_path, intermediates_path, videonames, traintest_parts, indices, feat_types, feats_path, \
pca_reduction=False, treelike=True, clusters_path=None, verbose=False):
try:
makedirs(feats_path)
except OSError:
pass
for k, part in enumerate(traintest_parts):
# cach'd pca and gmm
for j, feat_t in enumerate(feat_types):
try:
makedirs(join(feats_path, feat_t + '-' + str(k)))
except OSError:
pass
cache = None
# process videos
total = len(videonames)
for i in indices:
# FV computed for all feature types? see the last in INTERNAL_PARAMETERS['feature_types']
all_done = np.all([isfile(join(feats_path, feat_t + '-' + str(k), videonames[i] + '.pkl'))
for feat_t in feat_types])
if all_done:
if verbose:
print('[_compute_vd_descriptors] %s -> OK' % videonames[i])
continue
if cache is None:
cache = dict()
for j, feat_t in enumerate(feat_types):
with open(join(intermediates_path, 'gmm' + ('_pca-' if pca_reduction else '-') + feat_t + '-' + str(k) + '.pkl'), 'rb') as f:
cache[feat_t] = cPickle.load(f)
start_time = time.time()
# object features used for the per-frame FV representation computation (cach'd)
with open(join(tracklets_path, 'obj', videonames[i] + '.pkl'), 'rb') as f:
obj = cPickle.load(f)
with open(join(clusters_path, videonames[i] + '.pkl'), 'rb') as f:
clusters = cPickle.load(f)
for j, feat_t in enumerate(feat_types):
if isfile(join(feats_path, feat_t + '-' + str(k), videonames[i] + '.pkl')):
continue
# load video tracklets' feature
with open(join(tracklets_path, feat_t, videonames[i] + '.pkl'), 'rb') as f:
d = cPickle.load(f)
if feat_t == 'trj': # (special case)
d = convert_positions_to_displacements(d)
if feat_t == 'mbh':
dx = preprocessing.normalize(d[:,:d.shape[1]/2], norm='l1', axis=1)
dy = preprocessing.normalize(d[:,d.shape[1]/2:], norm='l1', axis=1)
d = np.hstack((dx,dy))
else:
d = preprocessing.normalize(d, norm='l1', axis=1)
d = rootSIFT(d)
if pca_reduction:
d = cache[feat_t]['pca'].transform(d) # reduce dimensionality
d = np.ascontiguousarray(d, dtype=np.float32) # required in many of Yael functions
output_filepath = join(feats_path, feat_t + '-' + str(k), videonames[i] + '.pkl')
# compute FV of the video
if not treelike:
# (in a per-frame representation)
fids = np.unique(obj[:,0])
V = [] # row-wise fisher vectors (matrix)
for f in fids:
tmp = d[np.where(obj[:,0] == f)[0],:] # hopefully this is contiguous if d already was
fv = ynumpy.fisher(cache[feat_t]['gmm'], tmp, include=INTERNAL_PARAMETERS['fv_repr_feats']) # f-th frame fisher vec
V.append(fv) # no normalization or nothing (it's done when computing darwin)
vd = videodarwin.darwin(np.array(V))
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(v=vd), f)
else: # or separately the FVs of the tree nodes
vdtree = dict()
if len(clusters['tree']) == 1:
fids = np.unique(obj[:,0])
V = [ynumpy.fisher(cache[feat_t]['gmm'], d[np.where(obj[:,0] == f)[0],:], INTERNAL_PARAMETERS['fv_repr_feats'])
for f in fids]
vdtree[1] = videodarwin.darwin(np.array(V))
else:
T = reconstruct_tree_from_leafs(np.unique(clusters['int_paths']))
for parent_idx, children_inds in T.iteritems():
# (in a per-frame representation)
node_inds = np.where(np.any([clusters['int_paths'] == idx for idx in children_inds], axis=0))[0]
fids = np.unique(obj[node_inds,0])
V = []
for f in fids:
tmp = d[np.where(obj[node_inds,0] == f)[0],:]
fv = ynumpy.fisher(cache[feat_t]['gmm'], tmp, INTERNAL_PARAMETERS['fv_repr_feats'])
V.append(fv) # no normalization or nothing (it's done when computing darwin)
vdtree[parent_idx] = videodarwin.darwin(np.array(V))
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(tree=vdtree), f)
elapsed_time = time.time() - start_time
if verbose:
print('[_compute_vd_descriptors] %s -> DONE (in %.2f secs)' % (videonames[i], elapsed_time))
def _compute_vd_descriptors(tracklets_path, intermediates_path, videonames, traintest_parts, indices, feat_types, feats_path, \
pca_reduction=True, treelike=True, clusters_path=None):
if not exists(feats_path):
makedirs(feats_path)
for k, part in enumerate(traintest_parts):
# cach'd pca and gmm
cache = dict()
for j, feat_t in enumerate(feat_types):
if not exists(feats_path + feat_t + '-' + str(k)):
makedirs(feats_path + feat_t + '-' + str(k))
with open(intermediates_path + 'gmm' + ('_pca-' if pca_reduction else '-') + feat_t + '-' + str(k) + '.pkl', 'rb') as f:
cache[feat_t] = cPickle.load(f)
# process videos
total = len(videonames)
for i in indices:
# FV computed for all feature types? see the last in INTERNAL_PARAMETERS['feature_types']
output_filepath = join(feats_path, feat_types[-1] + '-' + str(k), videonames[i] + '.pkl')
if isfile(output_filepath):
# for j, feat_t in enumerate(feat_types):
# featnames.setdefault(feat_t, []).append(feats_path + feat_t + '/' + videonames[i] + '-fvtree.pkl')
print('%s -> OK' % output_filepath)
continue
start_time = time.time()
# object features used for the per-frame FV representation computation (cach'd)
with open(tracklets_path + 'obj/' + videonames[i] + '.pkl', 'rb') as f:
obj = cPickle.load(f)
with open(clusters_path + videonames[i] + '.pkl', 'rb') as f:
clusters = cPickle.load(f)
for j, feat_t in enumerate(feat_types):
# load video tracklets' feature
with open(tracklets_path + feat_t + '/' + videonames[i] + '.pkl', 'rb') as f:
d = cPickle.load(f)
if feat_t == 'trj': # (special case)
d = convert_positions_to_displacements(d)
# pre-processing
d = rootSIFT(preprocessing.normalize(d, norm='l1', axis=1)) # https://hal.inria.fr/hal-00873267v2/document
if pca_reduction:
d = cache[feat_t]['pca'].transform(d) # reduce dimensionality
d = np.ascontiguousarray(d, dtype=np.float32) # required in many of Yael functions
output_filepath = join(feats_path, feat_t + '-' + str(k), videonames[i] + '.pkl')
# compute FV of the video
if not treelike:
# (in a per-frame representation)
fids = np.unique(obj[:,0])
V = [] # row-wise fisher vectors (matrix)
for f in fids:
tmp = d[np.where(obj[:,0] == f)[0],:] # hopefully this is contiguous if d already was
fv = ynumpy.fisher(cache[feat_t]['gmm'], tmp, include=INTERNAL_PARAMETERS['fv_repr_feats']) # f-th frame fisher vec
V.append(fv) # no normalization or nothing (it's done when computing darwin)
vd = normalize(videodarwin.darwin(np.array(V)))
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(v=vd), f)
else: # or separately the FVs of the tree nodes
T = reconstruct_tree_from_leafs(np.unique(clusters['int_paths']))
vdtree = dict()
for parent_idx, children_inds in T.iteritems():
# (in a per-frame representation)
node_inds = np.where(np.any([clusters['int_paths'] == idx for idx in children_inds], axis=0))[0]
fids = np.unique(obj[node_inds,0])
# dim = INTERNAL_PARAMETERS['fv_gmm_k'] * len(INTERNAL_PARAMETERS['fv_repr_feats']) * d.shape[1]
V = []
for f in fids:
tmp = d[np.where(obj[node_inds,0] == f)[0],:]
fv = ynumpy.fisher(cache[feat_t]['gmm'], tmp, INTERNAL_PARAMETERS['fv_repr_feats'])
V.append(fv) # no normalization or nothing (it's done when computing darwin)
vdtree[parent_idx] = normalize(videodarwin.darwin(np.array(V)))
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(tree=vdtree), f)
elapsed_time = time.time() - start_time
print('%s -> DONE (in %.2f secs)' % (videonames[i], elapsed_time))
def _compute_vd_descriptors(tracklets_path, intermediates_path, videonames, traintest_parts, indices, feat_types, feats_path, \
pca_reduction=False, treelike=True, clusters_path=None, verbose=False):
try:
makedirs(feats_path)
except OSError:
pass
for k, part in enumerate(traintest_parts):
# cach'd pca and gmm
for j, feat_t in enumerate(feat_types):
try:
makedirs(join(feats_path, feat_t + '-' + str(k)))
except OSError:
pass
cache = None
# process videos
total = len(videonames)
for i in indices:
# FV computed for all feature types? see the last in INTERNAL_PARAMETERS['feature_types']
all_done = np.all([
isfile(
join(feats_path, feat_t + '-' + str(k),
videonames[i] + '.pkl')) for feat_t in feat_types
])
if all_done:
if verbose:
print('[_compute_vd_descriptors] %s -> OK' % videonames[i])
continue
if cache is None:
cache = dict()
for j, feat_t in enumerate(feat_types):
with open(
join(
intermediates_path,
'gmm' + ('_pca-' if pca_reduction else '-') +
feat_t + '-' + str(k) + '.pkl'), 'rb') as f:
cache[feat_t] = cPickle.load(f)
start_time = time.time()
# object features used for the per-frame FV representation computation (cach'd)
with open(join(tracklets_path, 'obj', videonames[i] + '.pkl'),
'rb') as f:
obj = cPickle.load(f)
with open(join(clusters_path, videonames[i] + '.pkl'), 'rb') as f:
clusters = cPickle.load(f)
for j, feat_t in enumerate(feat_types):
if isfile(
join(feats_path, feat_t + '-' + str(k),
videonames[i] + '.pkl')):
continue
# load video tracklets' feature
with open(join(tracklets_path, feat_t, videonames[i] + '.pkl'),
'rb') as f:
d = cPickle.load(f)
if feat_t == 'trj': # (special case)
d = convert_positions_to_displacements(d)
if feat_t == 'mbh':
dx = preprocessing.normalize(d[:, :d.shape[1] / 2],
norm='l1',
axis=1)
dy = preprocessing.normalize(d[:, d.shape[1] / 2:],
norm='l1',
axis=1)
d = np.hstack((dx, dy))
else:
d = preprocessing.normalize(d, norm='l1', axis=1)
d = rootSIFT(d)
if pca_reduction:
d = cache[feat_t]['pca'].transform(
d) # reduce dimensionality
d = np.ascontiguousarray(
d, dtype=np.float32) # required in many of Yael functions
output_filepath = join(feats_path, feat_t + '-' + str(k),
videonames[i] + '.pkl')
# compute FV of the video
if not treelike:
# (in a per-frame representation)
fids = np.unique(obj[:, 0])
V = [] # row-wise fisher vectors (matrix)
for f in fids:
tmp = d[np.where(
obj[:, 0] == f
)[0], :] # hopefully this is contiguous if d already was
fv = ynumpy.fisher(
cache[feat_t]['gmm'],
tmp,
include=INTERNAL_PARAMETERS['fv_repr_feats']
) # f-th frame fisher vec
V.append(
fv
) # no normalization or nothing (it's done when computing darwin)
vd = videodarwin.darwin(np.array(V))
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(v=vd), f)
else: # or separately the FVs of the tree nodes
vdtree = dict()
if len(clusters['tree']) == 1:
fids = np.unique(obj[:, 0])
V = [
ynumpy.fisher(cache[feat_t]['gmm'],
d[np.where(obj[:, 0] == f)[0], :],
INTERNAL_PARAMETERS['fv_repr_feats'])
for f in fids
]
vdtree[1] = videodarwin.darwin(np.array(V))
else:
T = reconstruct_tree_from_leafs(
np.unique(clusters['int_paths']))
for parent_idx, children_inds in T.iteritems():
# (in a per-frame representation)
node_inds = np.where(
np.any([
clusters['int_paths'] == idx
for idx in children_inds
],
axis=0))[0]
fids = np.unique(obj[node_inds, 0])
V = []
for f in fids:
tmp = d[np.where(obj[node_inds, 0] == f)[0], :]
fv = ynumpy.fisher(
cache[feat_t]['gmm'], tmp,
INTERNAL_PARAMETERS['fv_repr_feats'])
V.append(
fv
) # no normalization or nothing (it's done when computing darwin)
vdtree[parent_idx] = videodarwin.darwin(
np.array(V))
with open(output_filepath, 'wb') as f:
cPickle.dump(dict(tree=vdtree), f)
elapsed_time = time.time() - start_time
if verbose:
print('[_compute_vd_descriptors] %s -> DONE (in %.2f secs)' %
(videonames[i], elapsed_time))