def calc_feature_vector(path):
x_cells, y_cells = good_cells(path, 3, 3)
avg_hists, bin_edges, avg_magnitudes, variances, flow = VideoFeatures(
path).calc_window_features(x_cells, y_cells, None)
fv = np.concatenate(
(avg_hists.flatten(), avg_magnitudes.flatten(), variances.flatten()))
return fv
def predict(self, path):
x_cells, y_cells = good_cells(path, 3, 3)
avg_hists, bin_edges, avg_magnitudes, variances, flow = VideoFeatures(
path).calc_window_features(x_cells, y_cells, None)
fv = np.concatenate((avg_hists.flatten(), avg_magnitudes.flatten(),
variances.flatten()))
prediction = self.classifier.predict(self.pca.transform(fv))[0]
return prediction
def predict(self, path):
x_cells, y_cells = good_cells(path, 3, 3)
avg_hists, bin_edges, avg_magnitudes, variances, flow = VideoFeatures(
path).calc_window_features(x_cells, y_cells, None)
fv = np.concatenate((avg_hists.flatten(), avg_magnitudes.flatten(),
variances.flatten()))
prediction = self.classifier.predict(self.pca.transform(fv))[0]
return prediction
def realtime_predict(self, test_path, label_path, window_size, start, end,
output):
labels = self.read_labels(label_path)
x_cells, y_cells = good_cells(test_path, 3, 3)
v = cv2.VideoWriter()
dims = vid_dims(test_path)
codec = cv2.cv.CV_FOURCC('m', 'p', '4', 'v')
v.open(output, codec, 15, dims, 1)
predictions = []
true_labels = []
for pos, agg_features in enumerate(
VideoFeatures(test_path, persist=True).aggregate_features(
x_cells, y_cells, window_size)):
if window_size + pos < start:
continue
if window_size + pos > end:
break
avg_hists, bin_edges, avg_magnitudes, variances, flow = agg_features
fv = np.concatenate((avg_hists.flatten(), avg_magnitudes.flatten(),
variances.flatten()))
if self.probabalistic:
prediction = self.classifier.predict_proba(
self.pca.transform(fv))[0]
else:
prediction = self.classifier.predict(self.pca.transform(fv))[0]
vis = flow.show("Prediction",
flow=False,
text=prediction,
display=True)
flow.show("Flow", flow=True, display=True)
v.write(vis)
# prediction is centered on current_time-window_size/2
# and current_time=window_size+pos
# => prediction is for (window_size/2)+pos
predictions.append(prediction)
true_labels.append(labels[(window_size / 2) + pos])
print
if self.probabalistic:
self.roc_analysis(predictions, true_labels)
else:
confmat_labels = sorted(set(predictions + true_labels))
confmat = confusion_matrix(true_labels, predictions,
confmat_labels)
self.draw_confmat(confmat, confmat_labels)
print "Accuracy: ", accuracy_score(true_labels, predictions)
v.release()
def calc_feature_vector(path): x_cells, y_cells = good_cells(path, 3, 3) avg_hists, bin_edges, avg_magnitudes, variances, flow = VideoFeatures(path).calc_window_features(x_cells, y_cells, None) fv = np.concatenate((avg_hists.flatten(), avg_magnitudes.flatten(), variances.flatten())) return fv
from bokeh.glyphs import ImageRGBA
from actipy.video_features import VideoFeatures
from actipy.plan import good_cells, vid_dims
import hashlib
from sys import argv
import matplotlib.pyplot as plt
path = argv[1]
output_path = argv[2]
vid_width, vid_height = vid_dims(path)
x_cells, y_cells = good_cells(path, 15, 15)
print "Using a cell grid of: %dx%d" % (x_cells, y_cells)
avg_hists, bin_edges, avg_magnitudes, variances, flow = VideoFeatures(
path).calc_window_features(x_cells, y_cells)
def make_color(x, y):
# RGB
#return "#%02x%02x%02x" % (np.floor(x*250), np.floor(np.logspace(0,np.log10(250),251)[np.floor(y*250)]), 150)
return "#%02x%02x%02x" % (np.floor(x * 250), np.floor(y * 250), 150)
bins = len(bin_edges) - 1
angle = 2.0 * np.pi / bins
#angles = np.arange(bins)*angle
angles = np.roll(bin_edges[:-1], bins / 4)
vis_width = 1200
vis_height = int(vid_height * (float(vis_width) / vid_width))