def find_norm_parameters(database_dir):
'''
Find normalization parameters for images, tilt values and pan values from training subset and store them in a .csv
file in the dataset directory.
Arguments:
database_dir: Directory containing the preprocessed dataset.
'''
# Get image, tilt and pan arrays from the preprocessed dataset.
img, tilt, pan = array_from_csv(database_dir + 'train.csv', database_dir)
# Calculate normalization parameters for the images.
i_mean = np.mean(img / 255.0)
i_std = np.std(img / 255.0)
# Calculate normalization parameters for the tilt values.
t_mean = np.mean(tilt / 90.0)
t_std = np.std(tilt / 90.0)
# Calculate normalization parameters for the pan values.
p_mean = np.mean(pan / 90.0)
p_std = np.std(pan / 90.0)
# Store normalization values in a .csv file.
file = open(database_dir + 'norm.csv', 'w')
file.write('img_mean,img_std,t_mean,t_std,p_mean,p_std\n')
file.write('%f,%f,%f,%f,%f,%f\n' % (i_mean, i_std, t_mean, t_std, p_mean, p_std))
# Close .csv file.
file.close()
def store_dataset_arrays(database_dir):
'''
Store preprocessed dataset images as a collection of .npy files.
Arguments:
database_dir: Directory containing the preprocessed dataset.
'''
# Load dataset as Numpy arrays.
tr_img, tr_tilt, tr_pan = array_from_csv(database_dir + 'train.csv', database_dir)
v_img, v_tilt, v_pan = array_from_csv(database_dir + 'validation.csv', database_dir)
t_img, t_tilt, t_pan = array_from_csv(database_dir + 'test.csv', database_dir)
# Store images as .npy arrays.
np.save(database_dir + 'train_img.npy', tr_img)
np.save(database_dir + 'validation_img.npy', v_img)
np.save(database_dir + 'test_img.npy', t_img)
sec_df = pd.read_csv(csv_path_2)
# Sort both datasets.
ref_df.sort_values(by=['tilt', 'pan'], inplace=True)
sec_df.sort_values(by=['tilt', 'pan'], inplace=True)
# Create the pandas dataframe that will contain the labels of the pictures from the second dataset with a
# correspondence in the reference dataset.
destination_df = pd.DataFrame(columns=ref_df.columns.values)
# Load pictures and pose values from the second dataset.
img = array_from_csv(csv_path_2, source_dir_2)[0]
# List that will contain the pictures from the second dataset with a correspondence in the reference dataset.
destination_list = []
# Naively search matching pictures between the two datasets using pose values as criterion.
start = 0
for row in sec_df.itertuples():
found = False
start_old = start
for i in range(start, len(ref_df.index)):
from clean_utils import array_from_csv
import os
dataset_dir = 'clean/aflw_pointing04/'
source_csv_file = 'test.csv'
destination_csv_file = 'test_aflw.csv'
destination_npy_file = 'test_aflw_img.npy'
source_csv = dataset_dir + source_csv_file
destination_csv = dataset_dir + destination_csv_file
destination_npy = dataset_dir + destination_npy_file
df = pd.read_csv(source_csv)
file = open(destination_csv, 'w')
file.write('file,tilt,pan\n')
for index, row in df.iterrows():
if int(row['file'][4:-4]) < 44782:
file.write(row['file'] + "," + str(row['tilt']) + "," +
str(row['pan']) + "\n")
file.close()
img_array, _, _ = array_from_csv(destination_csv, dataset_dir)
np.save(destination_npy, img_array)
dense_layer_size = 512
# Normalization parameters.
mean = 0.407335
std = 0.236271
t_mean = -0.022308
t_std = 0.324841
p_mean = 0.000171
p_std = 0.518044
# Load image, tilt and pan arrays for the dataset.
img, tilt, pan = array_from_csv(dataset_csv, dataset_dir)
# Estimator model.
pose_estimator = mpatacchiola_generic(in_size_estimator, num_conv_blocks,
num_filters_start, num_dense_layers,
dense_layer_size)
pose_estimator.load_weights(estimator_path)
# Get score for the dataset (tilt, pan and global error).
pred = pose_estimator.predict((img / 255.0 - mean) / std)
mean_tilt_error = np.mean(np.abs(tilt -
((pred[:, 0] * t_std + t_mean) * 90.0)))
mean_pan_error = np.mean(np.abs(pan - ((pred[:, 1] * p_std + p_mean) * 90.0)))
sec_df = pd.read_csv(dockerface_poses_file)
# Sort both datasets.
ref_df.sort_values(by=['tilt', 'pan'], inplace=True)
sec_df.sort_values(by=['tilt', 'pan'], inplace=True)
# Create the pandas dataframe that will contain the labels of the pictures from the second dataset with a
# correspondence in the reference dataset.
destination_df = pd.DataFrame(columns=['file', 'tilt', 'pan'])
# Load pictures from the Dockerface detections.
img = array_from_csv(dockerface_poses_file, aflw_dir)[0]
# List that will contain the pictures from the Dockerface detections with a correspondence in the reference dataset.
destination_list = []
# Grayscale version (for testing with RealHePoNet).
destination_list_grayscale = []
# Naively search matching pictures between the two datasets using pose values as criterion.
start = 0
for row in sec_df.itertuples():
found = False