-
Notifications
You must be signed in to change notification settings - Fork 13
/
Copy pathgenerate_training_FT.py
345 lines (281 loc) · 17.2 KB
/
generate_training_FT.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
# !/usr/bin/env python
# title :generate_training_FT.py
# description :Script that processes FT videos and input data from EYEDIAP dataset and generates training and
# testing data files compatible with our gaze estimation network. Data is filtered according to
# specific criteria. Transformation matrices for normalized space are also computed here.
# author :Cristina Palmero
# date :01062017
# version :2.0
# usage : -
# notes : FT: floating target. This script needs EYEDIAP dataset and folder structure to work.
# Change directories accordingly!
# python_version :3.5.5
# ==============================================================================
my_path = "F:\\EYEDIAP"
EYEDIAP_path = "EYEDIAP"
import os
import sys
sys.path.append(os.path.join(my_path, EYEDIAP_path, 'Scripts'))
from EYEDIAP_misc import * # script provided by EYEDIAP
from EYEDIAP_utils import *
from data_utils import *
import numpy as np
from data_utils import get_normalized_data, write_vector_to_file
mod = 'FT' # target type
type = ['S', 'M'] # type of head movement
# world coordinates to camera coordinates (see EYEDIAP for details)
Rw = [[1, 0, 0], [0, -1, 0], [0, 0, -1]]
tw = [0, 0, 1]
Rw = np.array(Rw)
tw = np.array(tw).reshape((3, 1))
for t in type:
# Session selection (from EYEDIAP scripts)
sessions = []
for P in range(1, 17):
if P < 12:
Cs = ['A']
elif P < 14:
Cs = ['B']
else:
Cs = ['A', 'B']
for C in Cs:
sessions.append(getSessionIndex(P, C, mod, t))
# Output files (all files are indexed:
# Validity: stores whether an specific frame is valid in terms of head pose, screen coords/point of gaze,
# correct gaze, face detection and face angles
validity_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('validity_' + mod + '_' + t + '.txt')), 'w')
# Data: stores directories of valid frames
data_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('data_' + mod + '_' + t + '.txt')), 'w')
# GTV: ground truth 3D gaze unit vectors of valid frames, in head coordinate system (HCS)
gtv_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('gtv_' + mod + '_' + t + '.txt')), 'wb')
# GT: ground truth 2D gaze direction angles (yaw and pitch) of valid frames wrt HCS
gt_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('gt_' + mod + '_' + t + '.txt')), 'wb')
# GTV cam: ground truth 3D gaze unit vectors of valid frames, in camera coordinate system (CCS)
gtv_cam_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('gtv_cam_' + mod + '_' + t + '.txt')), 'wb')
# GT cam: ground truth 2D gaze direction angles (yaw and pitch) of valid frames wrt CCS
gt_cam_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('gt_cam_' + mod + '_' + t + '.txt')), 'wb')
# GTH cam: ground truth 2D head direction angles of valid frames wrt CCS
gth_cam_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('gth_cam_' + mod + '_' + t + '.txt')), 'wb')
# GTV cam: ground truth 3D head direction unit vectors of valid frames wrt CCS
gthv_cam_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('gthv_cam_' + mod + '_' + t + '.txt')), 'wb')
# Face features: contains all needed characteristics of face and transformation matrices needed to (pre)process
# valid frames. This file contains the following 81 elements:
# (followed by their respective notation in BMVC paper):
# 0 seq_num (sequence number); 1 bb (bounding box);
# 2 face patch warp matrix (W); 3 face patch normalization mat. (M); 4 face patch gaze normalization matrix (R);
# 5 left eye patch warp matrix (W); 6 left eye patch norm. mat. (M); 7 left eye patch gaze norm. mat. (R);
# 8 right eye patch warp matrix (W); 9 right eye patch norm. mat. (M); 10 right eye patch gaze norm. mat. (R);
# 11 face ROI size; 12 eye ROI size; 13 - 80 3d landmarks
facefeats_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('face_features_' + mod + '_' + t + '.txt')), 'w')
# Sequence: information about sequences. A sequence is defined as a series of consecutive valid frames.
# Format: sequence number; video code;init frame; end frame;
sequence_file = open(os.path.join(my_path, EYEDIAP_path, 'Annotations', ('sequence_' + mod + '_' + t + '.txt')), 'w')
frames_done = 0
seq_num_t = 0
for session_num in sessions:
print("frames_done", frames_done)
print("Session num: ", session_num)
session_str = get_session_string(session_num)
print("Session string: ", session_str)
# session_str = '1_A_FT_M'
mpath = os.path.join(my_path, EYEDIAP_path, 'Eyes2', session_str)
if not os.path.exists(mpath):
os.makedirs(mpath)
# EYEDIAP data path
head_track_file = os.path.join(my_path, EYEDIAP_path, 'Data', session_str, 'head_pose.txt')
ball_track_file = os.path.join(my_path, EYEDIAP_path, 'Data', session_str, 'ball_tracking.txt')
valid_file = os.path.join(my_path, EYEDIAP_path, 'Annotations', 'GazeState', 'GazeStateExport', 'Data',
session_str,
'gaze_state.txt')
eyeball_centers_file = os.path.join(my_path, EYEDIAP_path, 'Metadata', 'Participants',
'eyes_position_%d' % int(session_str.split('_')[0]) + '.txt')
vga_calibration = os.path.join(my_path, EYEDIAP_path, 'Data', session_str, 'rgb_vga_calibration.txt')
# Video frames data path
frames_path = os.path.join(EYEDIAP_path, 'Data', session_str, 'frames')
# Landmarks file path
landmarks_file = os.path.join('F:\\landmarks', ('result3D_' + session_str + '.txt'))
# If annotations on valid frames are available, then these are loaded and taken into account
valid_frames = None
if os.path.exists(valid_file):
valid_frames = np.genfromtxt(valid_file, delimiter='\t', dtype=None)
# Frames path
frames_path = os.path.join(my_path, EYEDIAP_path, 'Data', session_str, 'frames')
frames = [os.path.join(frames_path, file) for file in os.listdir(frames_path) if file.endswith(".bmp")]
# Read the files with the frame-by-frame tracking parameters
ball_track = read_ball_track_file(ball_track_file)
head_track = read_head_track_file(head_track_file)
# Read Bulat et al ICCV2017 landmarks
# Format: one row per frame, all landmark coordinates separated by ' '
# e.g. "x1 y1 z1 x2 y2 z2 ...."
with open(landmarks_file, 'rb') as infile:
siz = sum(1 for _ in infile)
landmarks = np.empty((siz, 68, 3))
with open(landmarks_file, 'rb') as infile:
for i, line in enumerate(infile):
myline = line.split(b' ')
for j in list(range(0, 68)):
landmarks[i, j, :] = myline[3 * j:3 * j + 3]
# Read calibration file
calib = readCalibrationFile(vga_calibration)
# The estimated eyeball centers with respect to the head coordinate system
eyeball_centers = np.loadtxt(eyeball_centers_file)
# Containers for valid frames
valid_head = np.zeros(ball_track.shape[0], dtype=np.bool)
valid_ball = np.zeros(ball_track.shape[0], dtype=np.bool)
valid_gaze = np.zeros(ball_track.shape[0], dtype=np.bool)
valid_face = np.zeros(ball_track.shape[0], dtype=np.bool)
valid_angles = np.zeros(ball_track.shape[0], dtype=np.bool)
valid = False
start_sequence = 0
end_sequence = 0
seq_num = 0
eyeball_centers_moved = eyeball_centers.astype(dtype=np.float32, copy=True)
eyeball_centers_moved[:, 1] = eyeball_centers_moved[:, 1] + 0.005
for frameIndex in range(len(head_track[0])):
print("...................")
print("Frame: ", frameIndex)
# Read the parameters for the current frame
ball_pos = ball_track[frameIndex, :].reshape(3, 1)
R = head_track[0][frameIndex, :, :]
T = head_track[1][frameIndex, :].reshape(3, 1)
# Validity criteria
valid_head[frameIndex] = np.sum(T) != 0.0
valid_ball[frameIndex] = np.sum(ball_pos) != 0
valid_gaze[frameIndex] = valid_frames[frameIndex][1] == (b'OK' or b'BK') # Valid if OK or BlinKing
valid_face[frameIndex] = np.sum(landmarks[frameIndex]) > 0.0
# WCS to HCS system
# Refer the ball center to the head coordinate system
ball_HCS = np.dot(R.transpose(), ball_pos) - np.dot(R.transpose(), T)
# Generate the ground truth gaze vectors
gaze_vectors = ball_HCS.reshape(1, 3) - eyeball_centers
gaze_vectors = gaze_vectors / (np.sqrt(np.sum(gaze_vectors ** 2, axis=1)).reshape(-1, 1))
gaze_vector = np.mean(gaze_vectors, axis=0) # we use the mean of the gaze vectors
gaze_vector = gaze_vector / np.sqrt(np.sum(gaze_vector ** 2))
gaze_angles = vector2angles(gaze_vector)
# Valid angles according to (approx) max field of view when fixating in object
valid_angles[frameIndex] = (-40 * np.pi / 180 < gaze_angles[0][0] < 40 * np.pi / 180) \
and (-30 * np.pi / 180 < gaze_angles[0][1] < 30 * np.pi / 180)
valid_before = valid
valid = valid_head[frameIndex] and valid_ball[frameIndex] and valid_gaze[frameIndex] \
and valid_face[frameIndex] and valid_angles[frameIndex]
print("Valid: ", valid)
validity_row = np.array((valid_head[frameIndex], valid_ball[frameIndex], valid_gaze[frameIndex],
valid_angles[frameIndex], valid_face[frameIndex]))
# Write validity file
print("{};{};{}".format(session_str, format(frameIndex, '05'), np.array2string(validity_row, separator=';')),
file=validity_file)
# Write sequence file
# sequence number; video code;init frame; end frame;
if not valid:
if valid_before:
end_sequence = frameIndex - 1
print("{};{};{};{}".format(session_str, format(seq_num, '05'), format(start_sequence, '05'),
format(end_sequence, '05')),
file=sequence_file)
seq_num = seq_num + 1
seq_num_t = seq_num_t + 1
continue
if not valid_before:
start_sequence = frameIndex
# WCS to CCS system
eyeball_centers_WCS = np.dot(R, eyeball_centers.transpose()) + T
eyeball_centers_moved_WCS = np.dot(R, eyeball_centers_moved.transpose()) + T
eyeball_centers_CCS = np.dot(Rw.transpose(), eyeball_centers_WCS) - np.dot(Rw.transpose(), tw)
eyeball_centers_moved_CCS = np.dot(Rw.transpose(), eyeball_centers_moved_WCS) - np.dot(Rw.transpose(), tw)
ball_CCS = np.dot(Rw.transpose(), ball_pos) - np.dot(Rw.transpose(), tw)
gaze_vectors_CCS = ball_CCS.reshape(1, 3) - eyeball_centers_CCS.transpose()
gaze_vectors_CCS = gaze_vectors_CCS / (np.sqrt(np.sum(gaze_vectors_CCS ** 2, axis=1)).reshape(-1, 1))
gaze_vector_CCS = np.mean(gaze_vectors_CCS, axis=0)
gaze_vector_CCS = gaze_vector_CCS / np.sqrt(np.sum(gaze_vector_CCS ** 2))
gaze_angles_CCS = vector2angles(gaze_vector_CCS)
# Head direction in CCS
aux_vector = np.zeros((3, 1), dtype=np.float32)
aux_vector[2, 0] = 1.0
head_vector_WCS = np.dot(R, aux_vector)
head_vector_CCS = np.dot(Rw.transpose(), head_vector_WCS)
head_vector_CCS = head_vector_CCS / np.sqrt(np.sum(head_vector_CCS ** 2))
head_angle_CCS = vector2angles(head_vector_CCS)
# Face normalization
# we assume that center of face is 10 cm away from the center of the face in Z
head = np.array([[0.0], [0.0], [0.1]]).reshape(1, 3)
head_center_WCS = np.dot(head, R.transpose()) + T.transpose()
head_center_CCS = np.dot(Rw.transpose(), head_center_WCS.transpose()) - np.dot(Rw.transpose(), tw)
mean_face = head_center_CCS * 1000
mean_face = mean_face.transpose()
face_roi_size = [250, 250]
face_patch_conv, face_patch_warp, face_patch_gaze = get_normalized_data(mean_face[0], R, face_roi_size,
calib)
eyeball_centers_CCS1000 = eyeball_centers_moved_CCS * 1000
eyes_roi_size = [70, 58]
leye_patch_conv, leye_patch_warp, leye_patch_gaze = get_normalized_data(eyeball_centers_CCS1000[:,0], R,
eyes_roi_size, calib)
reye_patch_conv, reye_patch_warp, reye_patch_gaze = get_normalized_data(eyeball_centers_CCS1000[:,1], R,
eyes_roi_size, calib)
# Face Bounding box
# Get max distance between landmarks
max_dist = -1
for l1 in landmarks[frameIndex, :, :2]:
for l2 in landmarks[frameIndex,:,:2]:
if l1 is not l2:
dist = np.linalg.norm(l1 - l2)
if dist > max_dist:
max_dist = dist
mean_landmarks = np.mean(landmarks[frameIndex,:,:2], axis=0)
bb_height = max_dist
bb_dims = np.empty([4,1])
bb_dims[0] = mean_landmarks[0] - bb_height/2 # x
bb_dims[1] = mean_landmarks[1] - bb_height/2 # y
bb_dims[2] = bb_dims[3] = bb_height
# Write annotations in files
np.savetxt(gtv_cam_file, gaze_vector_CCS.reshape((1, 3)), fmt='%1.10f', delimiter='\t', newline='\n')
np.savetxt(gt_cam_file, gaze_angles_CCS, fmt='%1.10f', delimiter='\t', newline='\n')
np.savetxt(gtv_file, gaze_vector.reshape((1, 3)), fmt='%1.10f', delimiter='\t', newline='\n')
np.savetxt(gt_file, gaze_angles, fmt='%1.10f', delimiter='\t', newline='\n')
np.savetxt(gthv_cam_file, head_vector_CCS.reshape((1, 3)), fmt='%1.10f', delimiter='\t', newline='\n')
np.savetxt(gth_cam_file, head_angle_CCS, fmt='%1.10f', delimiter='\t', newline='\n')
img_file = os.path.join(frames_path, format(frameIndex, '05') + '.bmp')
data_file.write(img_file + '\n')
# 0 seq_num; 1 bb;
# 2 face patch warp; 3 face patch conv; 4 face patch gaze;
# 5 leye patch warp; 6 leye patch conv; 7 leye patch gaze;
# 8 reye patch warp; 9 reye patch conv; 10 reye patch gaze;
# 11 face roi size; 12 eye roi size; 13 - 80 3d landmarks
facefeats_file.write(str(seq_num_t) + ';')
facefeats_file.write(str(bb_dims[0][0]) + ',' + str(bb_dims[1][0]) + ',' + str(bb_dims[2][0]) + ','
+ str(bb_dims[3][0]) + ';')
write_vector_to_file(facefeats_file, face_patch_warp.reshape(9,1))
write_vector_to_file(facefeats_file, face_patch_conv.reshape(9,1))
write_vector_to_file(facefeats_file, face_patch_gaze.reshape(9, 1))
write_vector_to_file(facefeats_file, leye_patch_warp.reshape(9, 1))
write_vector_to_file(facefeats_file, leye_patch_conv.reshape(9, 1))
write_vector_to_file(facefeats_file, leye_patch_gaze.reshape(9, 1))
write_vector_to_file(facefeats_file, reye_patch_warp.reshape(9, 1))
write_vector_to_file(facefeats_file, reye_patch_conv.reshape(9, 1))
write_vector_to_file(facefeats_file, reye_patch_gaze.reshape(9, 1))
facefeats_file.write(str(face_roi_size[0]) + ',' + str(face_roi_size[1]) + ';')
facefeats_file.write(str(eyes_roi_size[0]) + ',' + str(eyes_roi_size[1]) + ';')
for j in list(range(0, 68)):
if j < 67:
facefeats_file.write(str(landmarks[frameIndex, j, 0]) + ',' + str(landmarks[frameIndex, j, 1])
+ ',' + str(landmarks[frameIndex, j, 2]) + ';')
else:
facefeats_file.write(str(landmarks[frameIndex, j, 0]) + ',' + str(landmarks[frameIndex, j, 1])
+ ',' + str(landmarks[frameIndex, j, 2]))
facefeats_file.write('\n')
if (frameIndex + 1) == len(head_track[0]):
end_sequence = frameIndex
print("{};{};{};{}".format(session_str, format(seq_num, '05'), format(start_sequence, '05'),
format(end_sequence, '05')),
file=sequence_file)
seq_num = seq_num + 1
seq_num_t = seq_num_t + 1
data_file.close()
gt_file.close()
gtv_file.close()
gt_cam_file.close()
gtv_cam_file.close()
gth_cam_file.close()
gthv_cam_file.close()
validity_file.close()
sequence_file.close()
facefeats_file.close()