def __init__(self, path, persist=False):
self.path = path
self.persist = persist
# video length minus one because the first
# frame doesn't have an optical flow
self.progress = ProgressBar(plan.vid_length(path) - 1)
class VideoFeatures:
"""
Given a video file extracts features.
"""
def __init__(self, path, persist=False):
self.path = path
self.persist = persist
# video length minus one because the first
# frame doesn't have an optical flow
self.progress = ProgressBar(plan.vid_length(path) - 1)
def loadFlows(self):
flowPath = "models/flow_%s.npy" % (hashlib.md5(self.path).hexdigest(),)
return np.load(flowPath)
def saveFlows(self, data):
flowPath = "models/flow_%s.npy" % (hashlib.md5(self.path).hexdigest(),)
return np.save(flowPath, data)
def features(self, x_cells, y_cells):
"""
Extracts HooF and flow magnitude from every frame of a video.
"""
if self.persist:
try:
generator = (i for i in self.loadFlows())
except:
generator = OpticalFlow(self.path).farneback()
else:
generator = OpticalFlow(self.path).farneback()
flows = []
print "Extracting features for %s..." % (self.path,)
for pos, flow in enumerate(generator):
flows.append(flow)
off = OpticalFlowFeatures(flow)
hist, bin_edges = off.cell_hoof(8, x_cells, y_cells, True)
# +1 because this iteration complete
# +1 because 0-based array indexing
self.progress.animate(pos+2)
yield hist, bin_edges, off.magnitude(x_cells, y_cells), flow
print
if self.persist:
self.saveFlows(flows)
def aggregate_features(self, x_cells, y_cells, window_size=None):
hists = []
magnitudes = []
for pos, fv in enumerate(self.features(x_cells, y_cells)):
hist, bin_edges, magnitude, flow = fv
if window_size and len(hists) == window_size:
yield self.summarise_features(hists, magnitudes, bin_edges) + (flow,)
hists.pop(0)
magnitudes.pop(0)
hists.append(hist)
magnitudes.append(magnitude)
if not window_size:
yield self.summarise_features(hists, magnitudes, bin_edges) + (flow,)
def summarise_features(self, hists, magnitudes, bin_edges):
#hists = np.swapaxes(np.swapaxes(hists, 0, 1), 1, 2)
#magnitudes = np.swapaxes(np.swapaxes(magnitudes, 0, 1), 1, 2)
# Now calculate mean of HooF
#import pdb; pdb.set_trace()
avg_hists = np.nanmean(hists, 0)
# Mean of flow magnitudes
avg_magnitudes = np.nanmean(magnitudes, 0)
avg_magnitudes = (avg_magnitudes-np.nanmin(avg_magnitudes))/(
np.nanmax(avg_magnitudes)-np.nanmin(avg_magnitudes))
# Sum of the variances over time in each bin in each cell
variances = np.sum(np.nanvar(hists, 0), 2)
variances = (variances-np.nanmin(variances))/(
np.nanmax(variances)-np.nanmin(variances))
return avg_hists, bin_edges, avg_magnitudes, variances
def calc_window_features(self, x_cells, y_cells, end=None):
"""
Legacy
"""
for fv in self.aggregate_features(x_cells, y_cells, end):
return fv