def detect_evaluate_spots(model,imagedir,train_frames,metadatafile,clf,stamp_shape,pset,num_empties=None):
filters = fg.get_hierarchical_filterbanks(model['config']['model']['layers'])
metadata = tb.tabarray(SVfile=metadatafile)
#get labels for training objects
transform_config = {'transform_name':'translation','percentile':pset}
train_labels_inds = []
for cn,fr in train_frames:
inds = ((metadata['Frame'] == fr) & (metadata['clip_num'] == cn) & (metadata['ObjectType'] != 'DCR')).nonzero()[0]
#ind = inds[t['object_number']]
train_labels_inds.extend(inds)
train_labels = metadata[train_labels_inds]
#get stamps for training objects
train_points = []
train_points_labels = []
sizes = []
num_train = 0
for label in train_labels:
lbl = label['clip_num'] + '_' + str(label['Frame']) + '.jpg'
print(label)
framefile = os.path.join(imagedir,lbl)
im = get_image(framefile)
box = bbox.BoundingBox(xs = [label[xf] for xf in xfields],
ys = [label[yf] for yf in yfields])
stamp = bbox.stamp(im,box,stamp_shape=stamp_shape)[0]
if stamp is not None:
sizes.append(stamp.shape)
print(stamp.shape)
try:
features = get_features(model,filters,stamp)
except:
print(label,'is bad')
else:
num_train += 1
features = {'0':features}
feature_stats = transform_average(features,transform_config,model)
train_points.append(feature_stats)
train_points_labels.append(1)
num_empties = (num_empties is not None) or num_train
for ind in range(num_empties):
print('empty',ind)
im = get_random_empty_bbox(train_labels,sizes,imagedir)
try:
features = get_features(model,filters,im)
except:
print('empty', ind, 'is bad')
else:
features = {'0':features}
feature_stats = transform_average(features,transform_config,model)
train_points.append(feature_stats)
train_points_labels.append(0)
train_points = np.array(train_points)
train_points_labels = np.array(train_points_labels)
prediction = clf.predict(train_points)
return prediction,train_points_labels
def detect_evaluate(model,imagedir,clf,test_frames,slice=None):
filters = fg.get_hierarchical_filterbanks(model['config']['model']['layers'])
#extract features from test frames
predictions = []
for clip_num,frame in test_frames:
framefile = os.path.join(imagedir,clip_num + '_' + str(frame) + '.jpg')
im = get_image(framefile)
if slice is not None:
im = im[slice]
tf = get_features(model,filters,im)
test_points,sh = get_feature_points(tf)
res = clf.predict(test_points)
predictions.append(res)
return predictions
def detect_train(model,metadatafile,imagedir,train_frames,num_empties=None,regress=True,points=True,pset=None,stamp_shape=0):
if points is False:
assert regress == False, 'regress must be false if points is False'
assert pset is not None, 'pset must not be nont if points is False'
assert stamp_shape > 0
transform_config = {'transform_name':'translation','percentile':pset}
filters = fg.get_hierarchical_filterbanks(model['config']['model']['layers'])
metadata = tb.tabarray(SVfile=metadatafile)
#get labels for training objects
train_labels_inds = []
for cn,fr in train_frames:
inds = ((metadata['Frame'] == fr) & (metadata['clip_num'] == cn) & (metadata['ObjectType'] != 'DCR')).nonzero()[0]
#ind = inds[t['object_number']]
train_labels_inds.extend(inds)
train_labels = metadata[train_labels_inds]
#get stamps for training objects
train_points = []
train_points_labels = []
sizes = []
num_train = 0
for label in train_labels:
lbl = label['clip_num'] + '_' + str(label['Frame']) + '.jpg'
print(label)
framefile = os.path.join(imagedir,lbl)
im = get_image(framefile)
box = bbox.BoundingBox(xs = [label[xf] for xf in xfields],
ys = [label[yf] for yf in yfields])
stamp = bbox.stamp(im,box,stamp_shape=stamp_shape)[0]
if stamp is not None:
sizes.append(stamp.shape)
print(stamp.shape)
try:
features = get_features(model,filters,stamp)
except:
print(label,'is bad')
else:
num_train += 1
if points:
feature_points,sh = get_feature_points(features)
train_points.extend(feature_points)
vecs = get_positions((box.width,box.height),features,regress=regress)
train_points_labels.extend(vecs)
else:
features = {'0':features}
feature_stats = transform_average(features,transform_config,model)
train_points.append(feature_stats)
train_points_labels.append(1)
num_empties = (num_empties is not None) or num_train
for ind in range(num_empties):
print('empty',ind)
im = get_random_empty_bbox(train_labels,sizes,imagedir)
try:
features = get_features(model,filters,im)
except:
print('empty', ind, 'is bad')
else:
if points:
feature_points,sh = get_feature_points(features)
train_points.extend(feature_points)
if regress:
vecs = [(-100,-100) for ind in range(len(feature_points))]
else:
vecs = [0 for ind in range(len(feature_points))]
train_points_labels.extend(vecs)
else:
features = {'0':features}
feature_stats = transform_average(features,transform_config,model)
train_points.append(feature_stats)
train_points_labels.append(0)
train_points = np.array(train_points)
train_points_labels = np.array(train_points_labels)
#run regression
if regress:
clf = linear_model.LinearRegression()
else:
clf = svm.LinearSVC()
clf.fit(train_points,train_points_labels)
return clf
