def train_classifiers(train_collection, hd, file_ext, cls_type='svc'):
train_data, train_labels, label_names, files = fe.generate_dataset(train_collection, hd, file_ext, False)
clf_l = Classifier('svc')
clf_r = Classifier('svc')
clf_l.train(train_data[0], train_labels[0])
clf_r.train(train_data[1], train_labels[1])
return clf_l, clf_r, label_names
features = 'honv' train_data, train_labels, train_label_names, _ = fe.generate_dataset(train_collection, hd, 'bmp', features, summarize) test_data_1, test_labels_1, test_label_names_1, _ = fe.generate_dataset(test_collection_b, hd_kin, 'bin', features, summarize) # Still Hand - Fingers Down test_data_2, test_labels_2, test_label_names_2, _ = fe.generate_dataset(test_collection_c, hd_kin, 'bin', features, summarize) # Scale test_data_3, test_labels_3, test_label_names_3, _ = fe.generate_dataset(test_collection_d, hd_kin, 'bin', features, summarize) # Practice Piece assert train_label_names == test_label_names_1, 'Train and test dataset do not have the same labels' assert train_label_names == test_label_names_2, 'Train and test dataset do not have the same labels' assert train_label_names == test_label_names_3, 'Train and test dataset do not have the same labels' # Test with SVC # left hand cls_svc = Classifier(cls_type='svc') cls_svc.train(train_data[0], train_labels[0]) score_svc_left_1 = cls_svc.score(test_data_1[0], test_labels_1[0]) score_svc_left_2 = cls_svc.score(test_data_2[0], test_labels_2[0]) score_svc_left_3 = cls_svc.score(test_data_3[0], test_labels_3[0]) # right hand cls_svc = Classifier(cls_type='svc') cls_svc.train(train_data[1], train_labels[1]) score_svc_right_1 = cls_svc.score(test_data_1[1], test_labels_1[1]) score_svc_right_2 = cls_svc.score(test_data_2[1], test_labels_2[1]) score_svc_right_3 = cls_svc.score(test_data_3[1], test_labels_3[1])
collection_3 = 'data/cv_all_3.col'
collection_4 = 'data/cv_all_combined.col'
hd_rs = DetectorFactory.get_hand_detector('realsense', 'grey')
hd_kin = DetectorFactory.get_hand_detector('kinect', 'depth')
summarize = False
features = 'hog'
data_1, labels_1, train_label_names_1, _ = fe.generate_dataset(collection_1, hd_rs, 'bmp', features, summarize)
data_2, labels_2, train_label_names_2, _ = fe.generate_dataset(collection_2, hd_kin, 'bin', features, summarize)
data_3, labels_3, train_label_names_3, _ = fe.generate_dataset(collection_3, hd_kin, 'bin', features, summarize)
data_4, labels_4, train_label_names_3, _ = fe.generate_dataset(collection_4, hd_kin, 'bin', features, summarize)
cv = 5
clf = Classifier('svc')
scores_l_1 = clf.cross_validation(data_1[0], labels_1[0], cv)
scores_r_1 = clf.cross_validation(data_1[1], labels_1[1], cv)
scores_l_2 = clf.cross_validation(data_2[0], labels_2[0], cv)
scores_r_2 = clf.cross_validation(data_2[1], labels_2[1], cv)
scores_l_3 = clf.cross_validation(data_3[0], labels_3[0], cv)
scores_r_3 = clf.cross_validation(data_3[1], labels_3[1], cv)
scores_l_4 = clf.cross_validation(data_4[0], labels_4[0], cv)
scores_r_4 = clf.cross_validation(data_4[1], labels_4[1], cv)
print "#### Cross Validation Results ####"
def error_analysis(collection, train_collection, background_model):
global l_avg_accum, r_avg_accum
train_data, train_labels, label_names, files = fe.generate_dataset(train_collection, background_model)
clf_l = Classifier('svc')
clf_r = Classifier('svc')
clf_l.train(train_data[0], train_labels[0])
clf_r.train(train_data[1], train_labels[1])
l_confusion_matrix = np.zeros((len(label_names), len(label_names)))
r_confusion_matrix = np.zeros((len(label_names), len(label_names)))
with open(collection, 'r') as f_collection:
print
print
print '***********************************'
print 'Error Analysis:', collection
print '***********************************'
# print '\t%s\t\t\t\t%s\t%s\t%s' % ('File Path', 'L', 'R', 'Actual')
l_errors = 0
r_errors = 0
n_samples = 0
sample_size = 10
# open each video in the collection
for line in f_collection:
a_line = line.replace('\n','').split('\t') # remove line breaks and split at tab
folder = a_line[0] # first element is folder containing video images
label = a_line[1] # next element is the ground truth label
# initialize hand detector
if label is '2':
hd = HandDetector('data/background_model_error2', folder) ### UPDATE: This is a hack for fixing some issues with error 2: will update after ICMC paper
else:
hd = HandDetector(background_model, folder)
right_sample = []
left_sample = []
i = 0
for left_hand, right_hand, filepath, img in hd.hand_generator():
if i<sample_size:
i += 1
right_sample.append(fe.extract_features(right_hand)) # extract the features
left_sample.append(fe.extract_features(left_hand))
else:
l_prediction = clf_l.majority_predict(np.array(left_sample))
r_prediction = clf_r.majority_predict(np.array(right_sample))
l_correct = l_prediction == int(label)
r_correct = r_prediction == int(label)
# if not (l_correct and r_correct):
# print '\t%s\t%s\t%s\t%s' % (filepath, l_prediction, r_prediction, label)
l_confusion_matrix[int(label)][l_prediction]+=1
r_confusion_matrix[int(label)][r_prediction]+=1
if not l_correct:
l_errors+=1
if not r_correct:
r_errors+=1
# Reset sample and populate with current hands
right_sample = []
left_sample = []
right_sample.append(fe.extract_features(right_hand))
left_sample.append(fe.extract_features(left_hand))
i = 1
n_samples+=1
print
print 'Error Rate:', collection
print 'Left:', (l_errors / float(n_samples))
print 'Right:', (r_errors / float(n_samples))
print
print
l_avg_accum += (l_errors / float(n_samples))
r_avg_accum += (r_errors / float(n_samples))
