def gen_gmms(base_dir='/home/econser/research/irsg_psu_pdx/',
train_file='personwearingglasses_fnames_train.txt',
n_components=3):
import irsg_utils as iu
data_dir = base_dir + 'data/'
image_dir = data_dir + 'PersonWearingGlasses/PersonWearingGlassesTrain/'
tt_split_dict = get_training_data(data_dir, train_file)
relationship_dict = get_pwg_annotations(image_dir, tt_split_dict)
wearing = relationship_dict['train']['wearing']
worn_by = relationship_dict['train']['worn_by']
gmms = {}
sk_gmm, platt = train_gmm(wearing, worn_by, n_components)
gmms['wearing'] = iu.RelationshipParameters(platt[0], platt[1],
sk_gmm.weights_, sk_gmm.means_,
sk_gmm.covariances_, sk_gmm)
sk_gmm, platt = train_gmm(worn_by, wearing, n_components)
gmms['worn_by'] = iu.RelationshipParameters(platt[0], platt[1],
sk_gmm.weights_, sk_gmm.means_,
sk_gmm.covariances_, sk_gmm)
return gmms
def gen_gmms(base_dir = '/home/econser/School/research/', train_file = 'dogwalkingtest_fnames_train.txt', test_file = 'dogwalkingtest_fnames_test.txt', n_components=3):
import irsg_utils as iu
data_dir = base_dir + 'data/'
image_dir = data_dir + 'dog_walking/'
tt_split_dict = get_train_test_splits(data_dir, train_file, test_file)
relationship_dict = get_dw_annotations(image_dir, tt_split_dict)
holding = relationship_dict['train']['holding']
attached_to = relationship_dict['train']['attached_to']
is_walking = relationship_dict['train']['is_walking']
walked_by = relationship_dict['train']['walked_by']
gmms = {}
sk_gmm, platt = train_gmm(holding, attached_to, n_components)
gmms['holding'] = iu.RelationshipParameters(platt[0], platt[1], sk_gmm.weights_, sk_gmm.means_, sk_gmm.covariances_, sk_gmm)
sk_gmm, platt = train_gmm(attached_to, holding, n_components)
gmms['attached_to'] = iu.RelationshipParameters(platt[0], platt[1], sk_gmm.weights_, sk_gmm.means_, sk_gmm.covariances_, sk_gmm)
sk_gmm, platt = train_gmm(is_walking, walked_by, n_components)
gmms['is_walking'] = iu.RelationshipParameters(platt[0], platt[1], sk_gmm.weights_, sk_gmm.means_, sk_gmm.covariances_, sk_gmm)
sk_gmm, platt = train_gmm(walked_by, is_walking, n_components)
gmms['walked_by'] = iu.RelationshipParameters(platt[0], platt[1], sk_gmm.weights_, sk_gmm.means_, sk_gmm.covariances_, sk_gmm)
return gmms
def gen_gmms(base_dir='/home/econser/research/',
train_file='pingpong_fnames_train.txt',
test_file='pingpong_fnames_test.txt',
n_components=3):
import irsg_utils as iu
data_dir = base_dir + 'data/'
image_dir = data_dir + 'PingPong/'
tt_split_dict = get_train_test_splits(data_dir, train_file, test_file)
relationship_dict = get_annotations(image_dir, tt_split_dict)
train_rel = []
train_rel.append(relationship_dict['train'][rel_str[0]])
train_rel.append(relationship_dict['train'][rel_str[1]])
train_rel.append(relationship_dict['train'][rel_str[2]])
neg_rel = []
neg_rel.append(relationship_dict['neg_train'][rel_str[0]])
neg_rel.append(relationship_dict['neg_train'][rel_str[1]])
neg_rel.append(relationship_dict['neg_train'][rel_str[2]])
gmms = {}
sk_gmm, platt = train_gmm(train_rel[0], neg_rel[0], n_components)
gmms[rel_str[0]] = iu.RelationshipParameters(platt[0], platt[1],
sk_gmm.weights_,
sk_gmm.means_,
sk_gmm.covariances_, sk_gmm)
sk_gmm, platt = train_gmm(train_rel[1], neg_rel[1], n_components)
gmms[rel_str[1]] = iu.RelationshipParameters(platt[0], platt[1],
sk_gmm.weights_,
sk_gmm.means_,
sk_gmm.covariances_, sk_gmm)
sk_gmm, platt = train_gmm(train_rel[2], neg_rel[2], n_components)
gmms[rel_str[2]] = iu.RelationshipParameters(platt[0], platt[1],
sk_gmm.weights_,
sk_gmm.means_,
sk_gmm.covariances_, sk_gmm)
return gmms
def gen_gmms(situation_cfg, training_fnames, cal_method, n_components=3):
# pull important info from the situation cfg
f_cfg = open(situation_cfg, 'rb')
j_cfg = j.load(f_cfg)
f_cfg.close()
situation_name = j_cfg['situation_name']
gmms = {}
relationships = j_cfg['relationships']
for rel_data in relationships:
rel_name = rel_data['name']
annotation_dir = rel_data['annotation_dir']
print('Creating model for "{}" relationship:'.format(rel_name))
training_set = training_fnames[rel_name]
relationship_dict = get_annotations(situation_name, annotation_dir,
training_set, rel_data)
pos_bboxes = relationship_dict['pos_bboxes']
print(' Found {} positive annotations'.format(len(pos_bboxes)))
gmm = train_gmm(pos_bboxes, n_components)
if cal_method == 'sample':
neg_bboxes = gen_neg_bboxes(gmm,
pos_bboxes,
neg_size=len(pos_bboxes))
else:
neg_bboxes = get_neg_bboxes(relationship_dict, rel_name)
cal_params = calibrate(gmm, pos_bboxes, neg_bboxes, title=rel_name)
gmms[rel_name] = iu.RelationshipParameters(cal_params[0],
cal_params[1], gmm.weights_,
gmm.means_,
gmm.covariances_, gmm)
return gmms