def train(train_indexes, BG_img, params):
folder = params["folder"]
marginX = params["marginX"]
marginY = params["marginY"]
neg_weight = params["neg_weight"]
method = params["method"]
feature = params["feature"]
train_feature_count = 0
train_features, train_labels = [], []
print "Extracting positive training features..."
# Read positive train features
for i in tqdm(range(len(train_indexes))):
img = img_read(folder, train_indexes[i])
#motion_img = read_motion_image(folder, train_indexes[i], BG_img)
height, width = img.shape
bboxes = add_bbox_margin(read_bboxes(folder, train_indexes[i]),
marginX, marginY, height, width)
for j in bboxes:
img_cut = img[j[0]:j[1], j[2]:j[3]]
#motion_img_cut = motion_img[j[0]:j[1], j[2]:j[3]]
train_feature_count += 1
#train_features.append(extract(img_cut, motion_img_cut, method, feature))
train_features.append(extract(img_cut, None, method, feature))
train_labels.append(1)
print "Positive training features are extracted."
print "Extracting negative training features..."
pos_train_feature_count = train_feature_count
# Read negative train features
for j in tqdm(range(pos_train_feature_count * neg_weight)):
i = sample(train_indexes, 1)[0]
img = img_read(folder, i)
height, width = img.shape
bboxes = add_bbox_margin(read_bboxes(folder, i), marginX, marginY,
height, width)
neg_bb = rand_bbox(bboxes, height, width)
if overlaps(neg_bb, bboxes) != -1:
continue
#motion_img = read_motion_image(folder, i, BG_img)
img_cut = img[neg_bb[0]:neg_bb[1], neg_bb[2]:neg_bb[3]]
#motion_img_cut = motion_img[neg_bb[0]:neg_bb[1], neg_bb[2]:neg_bb[3]]
train_feature_count += 1
#train_features.append(extract(img_cut, motion_img_cut, method, feature))
train_features.append(extract(img_cut, None, method, feature))
train_labels.append(-1)
print "Negative training features are extracted."
return train_features, train_labels, train_feature_count
def test(test_indexes, BG_img, params):
folder = params["folder"]
marginX = params["marginX"]
marginY = params["marginY"]
neg_weight = params["neg_weight"]
method = params["method"]
feature = params["feature"]
test_features, test_labels = [], []
test_feature_count = 0
print "Extracting positive test features..."
# Read positive test examples
for i in tqdm(range(len(test_indexes))):
img = img_read(folder, test_indexes[i])
#motion_img = read_motion_image(folder, test_indexes[i], BG_img)
height, width = img.shape
bboxes = add_bbox_margin(read_bboxes(folder, test_indexes[i]), marginX, marginY, height, width)
for j in bboxes:
img_cut = img[j[0]:j[1], j[2]:j[3]]
#motion_img_cut = motion_img[j[0]:j[1], j[2]:j[3]]
test_feature_count += 1
#test_features.append(extract(img_cut, motion_img_cut, method, feature))
test_features.append(extract(img_cut, None, method, feature))
test_labels.append(1)
pos_test_feature_count = test_feature_count
print "Positive test features are extracted."
print "Extracting negative test features..."
# Read negative test examples
for j in tqdm(range(pos_test_feature_count*neg_weight)):
i = sample(test_indexes, 1)[0]
img = img_read(folder, i)
height, width = img.shape
bboxes = add_bbox_margin(read_bboxes(folder, i), marginX, marginY, height, width)
neg_bb = rand_bbox(bboxes, height, width);
if overlaps(neg_bb, bboxes) != -1:
continue
#motion_img = read_motion_image(folder, i, BG_img)
img_cut = img[neg_bb[0]:neg_bb[1], neg_bb[2]:neg_bb[3]]
#motion_img_cut = motion_img[neg_bb[0]:neg_bb[1], neg_bb[2]:neg_bb[3]]
test_feature_count += 1
#test_features.append(extract(img_cut, motion_img_cut, method, feature))
test_features.append(extract(img_cut, None, method, feature))
test_labels.append(-1)
print "Negative test features are extracted."
return test_features, test_labels, test_feature_count
def bootstrap(bootstrap_indexes, BG_img, params, trf, trl, trfc, svm):
folder = params["folder"]
marginX = params["marginX"]
marginY = params["marginY"]
neg_weight = params["neg_weight"]
method = params["method"]
feature = params["feature"]
train_features = trf
train_labels = trl
train_feature_count = trfc
print "Starting bootstrapping..."
# Bootstrapping
for i in tqdm(range(len(bootstrap_indexes))):
img = img_read(folder, bootstrap_indexes[i])
#motion_img = read_motion_image(folder, bootstrap_indexes[i], BG_img)
bboxes = read_bboxes(folder, bootstrap_indexes[i])
#detections = detect_vehicles(img, motion_img, svm, params)
detections = detect_vehicles(img, None, svm, params)
hard_negatives = []
for j in detections:
if overlaps(j, bboxes) == -1:
hard_negatives.append(j)
height, width = img.shape
hard_negatives = add_bbox_margin(hard_negatives, marginX, marginY,
height, width)
for j in hard_negatives:
img_cut = img[j[0]:j[1], j[2]:j[3]]
#motion_img_cut = motion_img[j[0]:j[1], j[2]:j[3]]
train_feature_count += 1
#train_features.append(extract(img_cut, motion_img_cut, method, feature))
train_features.append(extract(img_cut, None, method, feature))
train_labels.append(-1)
print "Bootstrap finished."
return train_features, train_labels
def bootstrap(bootstrap_indexes, BG_img, params, trf, trl, trfc, svm):
folder = params["folder"]
marginX = params["marginX"]
marginY = params["marginY"]
neg_weight = params["neg_weight"]
method = params["method"]
feature = params["feature"]
train_features = trf
train_labels = trl
train_feature_count = trfc
print "Starting bootstrapping..."
# Bootstrapping
for i in tqdm(range(len(bootstrap_indexes))):
img = img_read(folder, bootstrap_indexes[i])
#motion_img = read_motion_image(folder, bootstrap_indexes[i], BG_img)
bboxes = read_bboxes(folder, bootstrap_indexes[i])
#detections = detect_vehicles(img, motion_img, svm, params)
detections = detect_vehicles(img, None, svm, params)
hard_negatives = []
for j in detections:
if overlaps(j, bboxes) == -1:
hard_negatives.append(j)
height, width = img.shape
hard_negatives = add_bbox_margin(hard_negatives, marginX, marginY, height, width)
for j in hard_negatives:
img_cut = img[j[0]:j[1], j[2]:j[3]]
#motion_img_cut = motion_img[j[0]:j[1], j[2]:j[3]]
train_feature_count += 1
#train_features.append(extract(img_cut, motion_img_cut, method, feature))
train_features.append(extract(img_cut, None, method, feature))
train_labels.append(-1)
print "Bootstrap finished."
return train_features, train_labels
