def main():
# Load parameters
params = get_args()
mouse_prec = params['mouse_prec']
mouse_speed = params['mouse_speed']
mouse = MouseController(mouse_prec, mouse_speed)
models = load_models(params)
# Load input feed
input_type = params['input_type']
if input_type=='cam':
input_file = None
else:
input_file = params['input_file_path']
feed=InputFeeder(input_type=input_type, input_file=input_file)
feed.load_data()
for batch in feed.next_batch():
if batch is not None:
image, pos = main_loop(batch, models)
cv2.imshow('frame', image)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
mouse.move(pos[0], pos[1])
# break
else:
break
feed.close()
def run_test(args):
mc = MouseController('medium', 'fast')
model_face = Face_Detector()
model_face.load_model(args.model_fd, args.device, CPU_EXTENSION)
model_pose = Pose_Estimator()
model_pose.load_model(args.model_pe, args.device, CPU_EXTENSION)
model_landmark = Facial_Landmarks()
model_landmark.load_model(args.model_fl, args.device, CPU_EXTENSION)
model_gaze = Gaze_Estimator()
model_gaze.load_model(args.model_ge, args.device, CPU_EXTENSION)
frame = cv2.imread(args.input)
crop_face, face_count, points = model_face.predict(frame, args.thres_fd)
if (face_count == 0):
print('no face is detected')
angles = model_pose.predict(frame, crop_face)
left_eye, right_eye, eye_points = model_landmark.predict(
frame, crop_face, points)
mx, my = model_gaze.predict(frame, left_eye, right_eye, angles, eye_points)
cv2.imwrite('images/ne.jpg', frame)
mc.move(mx, my)
def __init__(self,
device='CPU',
mouse_con=False,
face_dec=None,
fac_land=None,
head_pose=None,
gaze=None,
show_video=False,
save_video=False):
'''
all models should be put in here
'''
if face_dec and fac_land and head_pose and gaze:
self.face_dec, self.fac_land, self.head_pose, self.gaze = FaceDetectionModel(
face_dec, device=device), FacialLandmarksDetection(
fac_land, device=device), Head_Pose_Estimation(
head_pose,
device=device), Gaze_Estimation(gaze, device=device)
self.face_dec.load_model()
self.fac_land.load_model()
self.head_pose.load_model()
self.gaze.load_model()
else:
raise ValueError('Missing Arguments')
if mouse_con:
self.mouse_con = MouseController("low", "fast")
self.show_video, self.save_video = show_video, save_video
def __init__(self, args):
'''
This method instances variables for the Facial Landmarks Detection Model.
Args:
args = All arguments parsed by the arguments parser function
Return:
None
'''
init_start_time = time.time()
self.output_path = args.output_path
self.show_output = args.show_output
self.total_processing_time = 0
self.count_batch = 0
self.inference_speed = []
self.avg_inference_speed = 0
if args.all_devices != 'CPU':
args.face_device = args.all_devices
args.face_landmark_device = args.all_devices
args.head_pose_device = args.all_devices
args.gaze_device = args.all_devices
model_init_start = time.time()
self.face_model = FaceDetection(args.face_model, args.face_device,
args.face_device_ext,
args.face_prob_threshold)
self.landmarks_model = FacialLandmarksDetection(
args.face_landmark_model, args.face_landmark_device,
args.face_landmark_device_ext, args.face_landmark_prob_threshold)
self.head_pose_model = HeadPoseEstimation(
args.head_pose_model, args.head_pose_device,
args.head_pose_device_ext, args.head_pose_prob_threshold)
self.gaze_model = GazeEstimation(args.gaze_model, args.gaze_device,
args.gaze_device_ext,
args.gaze_prob_threshold)
self.model_init_time = time.time() - model_init_start
log.info('[ Main ] All required models initiallized')
self.mouse_control = MouseController(args.precision, args.speed)
log.info('[ Main ] Mouse controller successfully initialized')
self.input_feeder = InputFeeder(args.batch_size, args.input_type,
args.input_file)
log.info('[ Main ] Initialized input feeder')
model_load_start = time.time()
self.face_model.load_model()
self.landmarks_model.load_model()
self.head_pose_model.load_model()
self.gaze_model.load_model()
self.model_load_time = time.time() - model_load_start
self.app_init_time = time.time() - init_start_time
log.info('[ Main ] All moadels loaded to Inference Engine\n')
return None
def main():
"""
Load the network and parse the output.
:return: None
"""
# Grab command line args
mc = MouseController("high", "fast")
mc.move(randint(10, 109) / 100, -randint(10, 100) / 100)
def __init__(self, args):
# load the objects corresponding to the models
self.face_detection = Face_Detection(args.face_detection_model,
args.device, args.extensions,
args.perf_counts)
self.gaze_estimation = Gaze_Estimation(args.gaze_estimation_model,
args.device, args.extensions,
args.perf_counts)
self.head_pose_estimation = Head_Pose_Estimation(
args.head_pose_estimation_model, args.device, args.extensions,
args.perf_counts)
self.facial_landmarks_detection = Facial_Landmarks_Detection(
args.facial_landmarks_detection_model, args.device,
args.extensions, args.perf_counts)
start_models_load_time = time.time()
self.face_detection.load_model()
self.gaze_estimation.load_model()
self.head_pose_estimation.load_model()
self.facial_landmarks_detection.load_model()
logger = logging.getLogger()
input_T = args.input_type
input_F = args.input_file
if input_T.lower() == 'cam':
# open the video feed
self.feed = InputFeeder(args.input_type, args.input_file)
self.feed.load_data()
else:
if not os.path.isfile(input_F):
logger.error('Unable to find specified video file')
exit(1)
file_extension = input_F.split(".")[-1]
if (file_extension in ['jpg', 'jpeg', 'bmp']):
self.feed = InputFeeder(args.input_type, args.input_file)
self.feed.load_data()
elif (file_extension in ['avi', 'mp4']):
self.feed = InputFeeder(args.input_type, args.input_file)
self.feed.load_data()
else:
logger.error(
"Unsupported file Extension. Allowed ['jpg', 'jpeg', 'bmp', 'avi', 'mp4']"
)
exit(1)
print("Models total loading time :",
time.time() - start_models_load_time)
# init mouse controller
self.mouse_controller = MouseController('low', 'fast')
def main(args):
mouse_controller = MouseController('medium', 'fast')
print("Model Loading..")
face_detection = Model_FaceDetection(args.face_detection, args.device)
face_landmark = Model_FacialLandmarksDetection(args.face_landmark, args.device)
head_pose = Model_HeadPoseEstimation(args.head_pose, args.device)
gaze_estimation = Model_GazeEstimation(args.gaze_estimation, args.device)
print("Model loaded successfully")
input_feeder = InputFeeder(input_type='video', input_file=args.input)
input_feeder.load_data()
face_detection.load_model()
head_pose.load_model()
face_landmark.load_model()
gaze_estimation.load_model()
for frame in input_feeder.next_batch():
try:
frame.shape
except Exception as err:
break
key = cv2.waitKey(60)
face,face_coord = face_detection.predict(frame.copy(), args.prob_threshold)
if type(face)==int:
print("Unable to detect the face.")
if key==27:
break
continue
headPose = head_pose.predict(face.copy())
left_eye, right_eye, eye_coord = face_landmark.predict(face.copy())
mouse_coord, gaze_vector = gaze_estimation.predict(left_eye, right_eye, headPose)
cv2.imshow('video',frame)
mouse_controller.move(mouse_coord[0], mouse_coord[1])
input_feeder.close()
cv2.destroyAllWindows()
def model_instants(args):
face_detection_instant = FaceDetectionModel(model_name=args.face_detection,
device=args.device,
threshold=args.prob_threshold,
extensions=args.cpu_extension)
head_pose_estimation_instant = HeadPoseEstimationModel(
model_name=args.head_pose_estimation,
device=args.device,
extensions=args.cpu_extension)
facial_landmarks_instant = FacialLandmarksDetectionModel(
model_name=args.facial_landmarks_detection,
device=args.device,
extensions=args.cpu_extension)
gaze_estimation_instant = GazeEstimationModel(
model_name=args.gaze_estimation,
device=args.device,
extensions=args.cpu_extension)
mouse_controller_instant = MouseController('medium', 'fast')
return face_detection_instant, head_pose_estimation_instant, facial_landmarks_instant, gaze_estimation_instant, mouse_controller_instant
def main(model_dir, device, precision, input_type, input_file, inspect):
mouse_controller = MouseController("medium", "fast")
input_feeder = InputFeeder(input_type=input_type, input_file=input_file)
input_feeder.load_data()
gaze_detect = GazeDetect(model_dir=model_dir, device=device, precision=precision)
gaze_detect.load_model()
for image in input_feeder.next_batch():
with Timer() as t:
outputs = gaze_detect.predict(image)
if outputs is not None:
angle_y_fc, angle_p_fc, angle_r_fc = outputs.reshape(3)
mouse_controller.move(-angle_y_fc, angle_p_fc)
print(
f"Mouse move x: {-angle_y_fc}, y: {angle_p_fc}, execution time: {t.elapsed}"
)
def __init__(self, args):
self.log_level = "INFO" if os.environ.get(
"LOGLEVEL") == "INFO" or args.verbose_stage else "WARNING"
log.basicConfig(level=self.log_level)
input_type = 'cam' if args.cam else 'video'
self.feed = InputFeeder(input_type, args.video)
if not self.feed.load_data():
raise Exception('Input valid image or video file')
fps, w, h = self.feed.get_props()
self.out_video = cv2.VideoWriter(args.out,
cv2.VideoWriter_fourcc(*'MJPG'), fps,
(w, h), True)
args.head_pose_model = os.path.join(
args.head_pose_model, args.precision,
os.path.basename(args.head_pose_model))
args.landmarks_model = os.path.join(
args.landmarks_model, args.precision,
os.path.basename(args.landmarks_model))
args.gaze_model = os.path.join(args.gaze_model, args.precision,
os.path.basename(args.gaze_model))
self.fd = FaceDetect(args.face_model, args.device, args.extension,
args.threshold)
self.fd.load_model()
self.fd.set_out_size(w, h)
self.hp = HeadPoseEstimate(args.head_pose_model, args.device,
args.extension, args.threshold)
self.hp.load_model()
self.fl = FacialLandMarkDetect(args.landmarks_model, args.device,
args.extension, args.threshold)
self.fl.load_model()
self.gz = GazeEstimate(args.gaze_model, args.device, args.extension,
args.threshold)
self.gz.load_model()
self.mc = MouseController()
self.verbose_stage = args.verbose_stage
def main():
"""
Load the network and parse the output.
:return: None
"""
# Grab command line args
args = build_argparser().parse_args()
start_time = time.time()
face_detector = FaceDetect(model_name=args.face, device=args.device, output=args.output)
face_detector.load_model()
print("Time taken to load face detection model (in seconds):", time.time()-start_time)
start_time = time.time()
eyes_detector = EyesDetect(model_name=args.eyes, device=args.device, output=args.output)
eyes_detector.load_model()
print("Time taken to load landmark detection model (in seconds):", time.time()-start_time)
start_time = time.time()
angle_detector = AngleDetect(model_name=args.angle, device=args.device)
angle_detector.load_model()
print("Time taken to load head pose estimation model (in seconds):", time.time()-start_time)
start_time = time.time()
gaze_detector = GazeDetect(model_name=args.gaze, device=args.device)
gaze_detector.load_model()
print("Time taken to load gaze estimation model (in seconds):", time.time()-start_time)
mouse_controller = MouseController('medium','medium')
feed=InputFeeder(input_type=args.video, input_file=args.input)
feed.load_data()
for batch in feed.next_batch():
if batch is None: # catch last frame
break
face = face_detector.predict(batch)
left_eye, right_eye = eyes_detector.predict(face)
angles = angle_detector.predict(face)
x, y = gaze_detector.predict(left_eye, right_eye, angles)
mouse_controller.move(x, y)
feed.close()
def __init__(self):
self.args = None
self.feed = None
self.face_detection_model = None
self.facial_landmark_detection_model = None
self.gaze_estimation_model = None
self.head_pose_estimation_model = None
self.frame = None
self.width = None
self.Height = None
self.mc = MouseController("high", "fast")
self.face_detection_load_time = 0
self.facial_landmark_detection_load_time = 0
self.gaze_estimation_load_time = 0
self.head_pose_estimation_load_time = 0
self.face_detection_infer_time = 0
self.facial_landmark_detection_infer_time = 0
self.gaze_estimation_infer_time = 0
self.head_pose_estimation_infer_time = 0
self.frames = 0
def main(args):
feed = InputFeeder(input_type=args.it, input_file=args.i)
face_model = FaceDetectionModel(args.fm, args.d, args.c, float(args.p))
face_model.load_model()
landmarks_model = LandmarksDetectionModel(args.lm, args.d, args.c)
landmarks_model.load_model()
headpose_model = HeadPoseDetectionModel(args.hpm, args.d, args.c)
headpose_model.load_model()
gaze_model = GazeEstimationModel(args.gem, args.d, args.c)
gaze_model.load_model()
mouse = MouseController("medium", "fast")
feed.load_data()
for batch in feed.next_batch():
# try:
cropped_face, coords, _ = face_model.predict(batch)
cv2.rectangle(batch, (coords[0], coords[1]), (coords[2], coords[3]),
(255, 0, 0), 2)
left_eye, right_eye, eyes_coords, _ = landmarks_model.predict(
cropped_face)
head_pose_angles, _ = headpose_model.predict(cropped_face)
x, y, z, _ = gaze_model.predict(left_eye, right_eye, head_pose_angles,
cropped_face, eyes_coords)
mouse.move(x, y)
cv2.imshow("img", batch)
if cv2.waitKey(25) & 0xFF == ord('q'):
break
# except:
# print("Frame without prediction. Error: ", sys.exc_info()[0])
# log.error(sys.exc_info()[0])
feed.close()
def main():
mouseController = MouseController()
with camera() as cam:
while True:
ok, read_ = cam.read()
if not ok:
break
frame = cv2.flip(read_, 1)
roi = frame[ROI_SLICE].copy()
key = cv2.waitKey(2)
if key == ord('r'):
print('Background reset')
setBackground(roi)
continue
elif key in (27, ord('q')):
print('Exit')
break
mask = getMask(roi)
contour = getHandContour(roi, mask)
if contour is None or cv2.contourArea(contour) < MIN_HAND_SIZE:
continue
bottommost, leftmost, topmost, _rightmost = getHandPosition(
contour)
drawHandPositions(roi, bottommost, leftmost, topmost)
angle = getAngle(bottommost, leftmost, topmost)
if angle is None:
continue
mouseController.performActions(roi, topmost, angle)
showResult(frame, roi)
def inference_frame(m1,m2,m3,m4,inF,args):
"""
Funtion to calculate frame count and infernece time.
"""
visualize = args.visualize
mc = MouseController('high','fast')
total = 0
fc = 0
inf_time = 0
for ret, frame in inF.next_batch():
if not ret:
break;
if frame is not None:
fc += 1
if fc%5 == 0:
cv2.imshow('video', cv2.resize(frame, (500, 500)))
key = cv2.waitKey(60)
start_inf = time.time()
face_crop, face_dim = m1.predict(frame.copy(), args.prob_threshold)
if type(face_crop) == int:
print("No face detected.")
if key == 27:
break
continue
hp_out = m2.predict(face_crop.copy())
le, re, eye_dim = m3.predict(face_crop.copy())
new_dim, gv = m4.predict(le, re, hp_out)
end_inf = time.time()
inf_time = inf_time + end_inf - start_inf
total = total + 1
visualization(visualize, frame, face_crop, face_dim, eye_dim, hp_out, gv, le, re)
if fc%5 == 0:
mc.move(new_dim[0], new_dim[1])
if key == 27:
break
return fc,inf_time
def start_pipeline(cla, codec):
"""
Initializes feeds inputs to models, moving the mouse cursor based on the final gaze estimation.
:param cla: Command line arguments for configuring the pipeline.
:param codec: Depending on the platform this is run on, OpenCV requires a codec to be specified. Supply it here.
:return: None
"""
preview_flags = cla.preview_flags
logger = logging.getLogger()
input_file_path = cla.input
if input_file_path.lower() == "cam":
in_feeder = InputFeeder("cam")
elif not os.path.isfile(input_file_path):
# top = os.path.dirname(os.path.realpath(__file__))
# walktree(top, visit_file)
logger.error("Cannot locate video file provided. Exiting..")
sys.exit(1)
else:
in_feeder = InputFeeder("video", input_file_path)
start_model_load_time = time.time()
fdm, fldm, hpem, gem = prep_models(cla)
total_model_load_time = time.time() - start_model_load_time
mc = None
if not cla.is_benchmark:
mc = MouseController('medium', 'fast')
in_feeder.load_data()
fps, total_inference_time, total_time = handle_input_feed(
logger, preview_flags, fdm, fldm, hpem, gem, mc, in_feeder,
cla.frame_out_rate, codec, cla.output_path)
with open(os.path.join(cla.output_path, 'stats.txt'), 'w') as f:
f.write("Total inference time, " + str(total_inference_time) + '\n')
f.write("FPS, " + str(fps) + '\n')
f.write("Total model load time, " + str(total_model_load_time) + '\n')
f.write("Total time, " + str(total_time) + '\n')
logger.error("Video stream ended...")
cv2.destroyAllWindows()
in_feeder.close()
def init_models(args):
global model_face, model_landmarks, model_hpose, model_gaze_estimation, mouse_controller
precision = args.precision
device = args.device
threshold = args.threshold
model_face = ModelFaceDetection(model_dir_face[precision], device, threshold)
model_landmarks = ModelFacialLandmarksDetection(model_dir_landmarks[precision], device, threshold)
model_hpose = ModelHeadPoseEstimation(model_dir_hpose[precision], device, threshold)
model_gaze_estimation = ModelGazeEstimation(model_dir_gaze[precision], device, threshold)
model_face.load_model()
model_landmarks.load_model()
model_hpose.load_model()
model_gaze_estimation.load_model()
mouse_controller = None
if args.mouse_precision in ['high', 'low', 'medium'] and args.mouse_speed in ['fast', 'slow', 'medium']:
mouse_controller = MouseController(args.mouse_precision, args.mouse_speed)
def setup(args):
global input_path, output_path, device, cpu_extension, prob_threshold, flags, mouse_controller, feeder, video_writer, model_dict, model_loading_total_time
model_args = [
args.face_detection_model,
args.facial_landmarks_detection_model,
args.head_pose_estimation_model,
args.gaze_estimation_model,
]
model_class = [
Model_FaceDetection,
Model_FacialLandMarkDetection,
Model_HeadPoseEstimation,
Model_GazeEstimation,
]
input_path = input_path_generator(args.input) if args.input != "CAM" else None
output_path = output_path_generator(args.output)
device = args.device
cpu_extension = args.cpu_extension
prob_threshold = args.prob_threshold
flags = args.flags
if not os.path.exists(output_path):
os.mkdir(output_path)
mouse_controller = MouseController("low", "fast")
if input_path:
if input_path.endswith(".jpg"):
feeder = InputFeeder("image", input_path)
else:
feeder = InputFeeder("video", input_path)
else:
feeder = InputFeeder("cam")
feeder.load_data()
fps = feeder.fps()
initial_w, initial_h, video_len = feeder.frame_initials_and_length()
video_writer = cv2.VideoWriter(
os.path.join(output_path, "output_video.mp4"),
cv2.VideoWriter_fourcc(*"avc1"),
fps / 10,
(initial_w, initial_h),
True,
)
model_dict, model_loading_total_time = generate_model_dict(model_args, model_class)
return
def init_model(args):
global face_model, landmark_model, head_pose_model, gaze_model, mouse_controller
device_name = args.device
prob_threshold = args.prob_threshold
# Initialize variables with the input arguments for easy access
model_path_dict = {
'FaceDetectionModel': args.faceDetectionModel,
'LandmarkRegressionModel': args.landmarkRegressionModel,
'HeadPoseEstimationModel': args.headPoseEstimationModel,
'GazeEstimationModel': args.gazeEstimationModel
}
# Instantiate model
face_model = Model_Face(model_path_dict['FaceDetectionModel'],
device_name,
threshold=prob_threshold)
landmark_model = Model_Landmark(model_path_dict['LandmarkRegressionModel'],
device_name,
threshold=prob_threshold)
head_pose_model = Model_Pose(model_path_dict['HeadPoseEstimationModel'],
device_name,
threshold=prob_threshold)
gaze_model = Model_Gaze(model_path_dict['GazeEstimationModel'],
device_name,
threshold=prob_threshold)
mouse_controller = MouseController('medium', 'fast')
# Load Models
face_model.load_model()
landmark_model.load_model()
head_pose_model.load_model()
gaze_model.load_model()
# Check extention of these unsupported layers
face_model.check_model()
landmark_model.check_model()
head_pose_model.check_model()
gaze_model.check_model()
def main():
logging.info("Parsing the arguments.")
args = get_parser().parse_args()
logging.info("Arguments parsed successfully. Now initialing feedreader.")
input_feeder = init_feeder(args)
logging.info("FeedReader initialized, loading the models.")
fd_model, fld_model, ge_model, hpe_model, total_model_load_time = load_all_models(
args)
mc = MouseController('medium', 'fast')
logging.info("Starting the workflow")
fps, total_inference_time, effective_fps = run_workflow(
fd_model, fld_model, ge_model, hpe_model, input_feeder, mc,
args.show_intermediate_visualization)
logging.debug("Writing the stats.")
with open(os.path.join(args.output_dir, 'stats.txt'), 'w') as f:
f.write(str(total_inference_time) + '\n')
f.write(str(fps) + '\n')
f.write(str(effective_fps) + '\n')
f.write(str(total_model_load_time) + '\n')
def gaze_pointer_controller(args, facedetector, facelm, headpose, gaze):
mouse_controller = MouseController(precision='high', speed='fast')
# Handle input type:
inference_time_face = []
inference_time_landmarks = []
inference_time_headpose = []
inference_time_gaze = []
if args.input != 'CAM':
try:
# It seems that OpenCV can use VideoCapture to treat videos and images:
input_stream = cv2.VideoCapture(args.input)
length = int(input_stream.get(cv2.CAP_PROP_FRAME_COUNT))
webcamera = False
# Check if input is an image or video file:
if length > 1:
single_image_mode = False
else:
single_image_mode = True
except:
print(
'Not supported image or video file format. Please pass a supported one.'
)
exit()
else:
input_stream = cv2.VideoCapture(0)
single_image_mode = False
webcamera = True
if not single_image_mode:
count = 0
while (input_stream.isOpened()):
# Read the next frame:
flag, frame = input_stream.read()
if not flag:
break
if count % args.frame_count == 0:
start = time.time()
if cv2.waitKey(1) & 0xFF == ord('q'):
break
# We get a detected face crop and its coordinates:
face_crop, detection = facedetector.get_face_crop(frame, args)
finish_face_detector_time = time.time()
face_detector_time = round(finish_face_detector_time - start,
5)
log.info("Face detection took {} seconds.".format(
face_detector_time))
inference_time_face.append(face_detector_time)
# Obtain eyes coordinates:
right_eye, left_eye = facelm.get_eyes_coordinates(face_crop)
# Obtain eyes crops:
right_eye_crop, left_eye_crop, right_eye_coords, left_eye_coords = utils.get_eyes_crops(
face_crop, right_eye, left_eye)
finish_eyes_coordinates = time.time()
eyes_detector_time = round(
finish_eyes_coordinates - finish_face_detector_time, 5)
log.info("Eyes detection took {} seconds.".format(
eyes_detector_time))
inference_time_landmarks.append(eyes_detector_time)
# Obtain headpose angles:
headpose_angles = headpose.get_headpose_angles(face_crop)
finish_headpose_angles = time.time()
headpose_detector_time = round(
finish_headpose_angles - finish_eyes_coordinates, 5)
log.info("Headpose angles detection took {} seconds.".format(
headpose_detector_time))
inference_time_headpose.append(headpose_detector_time)
# Obtain gaze vector and mouse movement values:
(x_movement, y_movement), gaze_vector = gaze.get_gaze(
right_eye_crop, left_eye_crop, headpose_angles)
finish_gaze_detection_time = time.time()
gaze_detector_time = round(
finish_gaze_detection_time - finish_headpose_angles, 5)
log.info("Gaze detection took {} seconds.".format(
gaze_detector_time))
inference_time_gaze.append(gaze_detector_time)
# Optional visualization configuration:
if args.view_face:
frame = cv2.rectangle(frame, (detection[0], detection[1]),
(detection[2], detection[3]),
color=(0, 255, 0),
thickness=5)
if args.view_eyes:
right_eye_coords = [
right_eye_coords[0] + detection[1],
right_eye_coords[1] + detection[1],
right_eye_coords[2] + detection[0],
right_eye_coords[3] + detection[0]
]
left_eye_coords = [
left_eye_coords[0] + detection[1],
left_eye_coords[1] + detection[1],
left_eye_coords[2] + detection[0],
left_eye_coords[3] + detection[0]
]
frame = cv2.rectangle(
frame, (right_eye_coords[2], right_eye_coords[1]),
(right_eye_coords[3], right_eye_coords[0]),
color=(255, 0, 0),
thickness=5)
frame = cv2.rectangle(
frame, (left_eye_coords[2], left_eye_coords[1]),
(left_eye_coords[3], left_eye_coords[0]),
color=(255, 0, 0),
thickness=5)
if args.view_gaze:
# Right eye:
x_r_eye = int(right_eye[0] * face_crop.shape[1] +
detection[0])
y_r_eye = int(right_eye[1] * face_crop.shape[0] +
detection[1])
x_r_shift, y_r_shift = int(x_r_eye + gaze_vector[0] *
100), int(y_r_eye -
gaze_vector[1] * 100)
# Left eye:
x_l_eye = int(left_eye[0] * face_crop.shape[1] +
detection[0])
y_l_eye = int(left_eye[1] * face_crop.shape[0] +
detection[1])
x_l_shift, y_l_shift = int(x_l_eye + gaze_vector[0] *
100), int(y_l_eye -
gaze_vector[1] * 100)
frame = cv2.arrowedLine(frame, (x_r_eye, y_r_eye),
(x_r_shift, y_r_shift),
(0, 0, 255), 2)
frame = cv2.arrowedLine(frame, (x_l_eye, y_l_eye),
(x_l_shift, y_l_shift),
(0, 0, 255), 2)
if args.view_headpose:
frame = cv2.putText(
frame, 'Yaw: ' + str(headpose_angles[0]) + ' ' +
'Pitch: ' + str(headpose_angles[1]) + ' ' + 'Roll: ' +
str(headpose_angles[2]), (15, 20),
cv2.FONT_HERSHEY_SIMPLEX, 0.65, (0, 0, 0), 2)
# Resizing window for visualization convenience:
cv2.namedWindow('Prueba', cv2.WINDOW_NORMAL)
cv2.resizeWindow('Prueba', 600, 400)
cv2.imshow('Prueba', frame)
mouse_controller.move(x_movement, y_movement)
count = count + 1
input_stream.release()
with open('times.csv', 'w', newline='') as csvfile:
writer = csv.writer(csvfile, delimiter=',')
writer.writerow([
'Face Detector', 'Eyes Detector', 'Headpose Detector',
'Gaze Detector'
])
for i in range(len(inference_time_face)):
writer.writerow([
inference_time_face[i], inference_time_landmarks[i],
inference_time_headpose[i], inference_time_gaze[i]
])
cv2.destroyAllWindows()
def main():
args = build_argparser().parse_args()
visual = args.visual_flag
log = logging.getLogger()
input_source = args.input_source
try:
video_path = args.input_path
except Exception as e:
video_path = None
feed = None
if input_source.lower() == 'cam':
feed = InputFeeder('cam')
elif input_source.lower() == 'video' and os.path.isfile(video_path):
feed = InputFeeder('video', video_path)
else:
log.error('Wrong input feed. (check the video path).')
exit(1)
fd = Model_Face(args.face_detection_model, args.device, args.extension)
hp = Model_HeadPose(args.head_pose_model, args.device, args.extension)
fl = Model_Faciallandmark(args.facial_landmarks_model, args.device,
args.extension)
ga = Model_Gaze(args.gaze_model, args.device, args.extension)
### You can specify the value of precision and speed directly.
## OR
## 'high'(100),'low'(1000),'medium','low-med' - precision
## 'fast'(1), 'slow'(10), 'medium', 'slow-med' - speed
# mouse = MouseController('low-med', 'slow-med')
mouse = MouseController(500, 4)
feed.load_data()
# load models
fd.load_model()
hp.load_model()
fl.load_model()
ga.load_model()
count = 0
for ret, frame in feed.next_batch():
if not ret:
break
count += 1
if count % 5 == 0:
cv2.imshow('video', cv2.resize(frame, (500, 500)))
key = cv2.waitKey(60)
frame_cp = frame.copy()
face, face_position = fd.predict(frame_cp, args.threshold)
if type(face) == int:
log.error('Prediction Error: Cant find face.')
if key == 27:
break
continue
face_cp = face.copy()
hp_output = hp.predict(face_cp)
left_eye, right_eye, facial = fl.predict(face_cp)
# print('left',left_eye,'\n','right',right_eye,'\n')
mouse_coord, gaze_vector = ga.predict(left_eye, right_eye, hp_output)
if (not len(visual) == 0):
visual_frame = frame.copy()
### Visual FLAGS
# face detection
if 'fd' in visual:
visual_frame = face
# Head pose
if 'hp' in visual:
cv2.putText(
visual_frame,
"Yaw: {:.2f} Pitch: {:.2f} Roll: {:.2f}".format(
hp_output[0], hp_output[1], hp_output[2]), (10, 20),
cv2.FONT_HERSHEY_COMPLEX, 0.3, (0, 255, 50), 1)
# Facial landmarks
if 'fl' in visual:
cv2.rectangle(face, (facial[0][0] - 10, facial[0][1] - 10),
(facial[0][2] + 10, facial[0][3] + 10),
(255, 0, 0), 3)
cv2.rectangle(face, (facial[1][0] - 10, facial[1][1] - 10),
(facial[1][2] + 10, facial[1][3] + 10),
(255, 0, 0), 3)
# Gaze estimation
if 'ga' in visual:
x, y, w = int(gaze_vector[0] * 12), int(gaze_vector[1] *
12), 160
le = cv2.line(left_eye.copy(), (x - w, y - w), (x + w, y + w),
(255, 255, 0), 2)
cv2.line(le, (x - w, y + w), (x + w, y - w), (255, 50, 150), 2)
re = cv2.line(right_eye.copy(), (x - w, y - w), (x + w, y + w),
(255, 255, 0), 2)
cv2.line(re, (x - w, y + w), (x + w, y - w), (255, 50, 150), 2)
face[facial[0][1]:facial[0][3], facial[0][0]:facial[0][2]] = le
face[facial[1][1]:facial[1][3], facial[1][0]:facial[1][2]] = re
cv2.namedWindow('Visualization', cv2.WINDOW_AUTOSIZE)
cv2.moveWindow('Visualization', 900, 900)
cv2.imshow('Visualization', cv2.resize(visual_frame, (500, 500)))
if args.visual_save.lower() == 'y':
if count % 10 == 0:
cv2.imwrite(str(count) + '_visual.jpg', visual_frame)
if count % 5 == 0:
mouse.move(mouse_coord[0], mouse_coord[1])
if key == 27:
break
log.error('INFO: Ended!')
cv2.destroyAllWindows()
feed.close()
def main():
args = get_args().parse_args()
path_filender = args.input
four_flags = args.flags_checker
loger = logging.getLogger()
feeder_in = None
out_path = args.out_path
if path_filender.lower() == "cam":
feeder_in = InputFeeder("cam")
else:
if not os.path.isfile(path_filender):
loger.error("The video was not found")
exit(1)
feeder_in = InputFeeder("video", path_filender)
model_locations = {
'FaceDetection': args.face_detection_model,
'HeadPoseEstimation': args.head_pose_estimation_model,
'FacialLandmarksDetection': args.facial_landmarks_detection_model,
'GazeEstimation': args.gaze_estimation_model
}
for key_name in model_locations.keys():
if not os.path.isfile(model_locations[key_name]):
loger.error("The system cannot find the " + key_name + " xml file")
exit(1)
dt = FaceDetection(model_locations['FaceDetection'], args.device,
args.cpu_extension)
pe = HeadPoseEstimation(model_locations['HeadPoseEstimation'], args.device,
args.cpu_extension)
ld = FacialLandmarksDetection(model_locations['FacialLandmarksDetection'],
args.device, args.cpu_extension)
ge = GazeEstimation(model_locations['GazeEstimation'], args.device,
args.cpu_extension)
cursor = MouseController('medium', 'fast')
feeder_in.load_data()
model_load_time_start = time.time()
dt.load_model()
pe.load_model()
ld.load_model()
ge.load_model()
total_load_time = time.time() - model_load_time_start
frame_counter = 0
inference_time_start = time.time()
for ret, frame in feeder_in.next_batch():
if not ret:
break
frame_counter = frame_counter + 1
if frame_counter % 1 == 0:
cv2.imshow('video', cv2.resize(frame, (600, 600)))
key = cv2.waitKey(60)
face_detected, coords_face = dt.predict(frame, args.p_th)
if type(face_detected) == int:
loger.error("The system cannot detect any face.")
if key == 27:
break
continue
head_pose_output = pe.predict(face_detected)
eye_left_detect, eye_right_detect, eye_coordinates_detect = ld.predict(
face_detected)
coordi_update_pointer, coordi_gaze = ge.predict(
eye_left_detect, eye_right_detect, head_pose_output)
if (not len(four_flags) == 0):
result_app = frame
if 'fad' in four_flags:
result_app = face_detected
if 'hpe' in four_flags:
cv2.putText(
result_app,
"HP Angles: YAW:{:.3f} * PITCH:{:.3f} * ROLL:{:.3f}".
format(head_pose_output[0], head_pose_output[1],
head_pose_output[2]), (5, 40),
cv2.FONT_HERSHEY_COMPLEX, 0.25, (153, 76, 0), 0)
if 'fld' in four_flags:
cv2.rectangle(face_detected,
(eye_coordinates_detect[0][0] - 4,
eye_coordinates_detect[0][1] - 4),
(eye_coordinates_detect[0][2] + 4,
eye_coordinates_detect[0][3] + 4),
(255, 255, 0), 4)
cv2.rectangle(face_detected,
(eye_coordinates_detect[1][0] - 4,
eye_coordinates_detect[1][1] - 4),
(eye_coordinates_detect[1][2] + 4,
eye_coordinates_detect[1][3] + 4),
(255, 255, 0), 4)
if 'gae' in four_flags:
x = int(coordi_gaze[0] * 2)
y = int(coordi_gaze[1] * 2)
w = 150
right_E = cv2.line(eye_right_detect, (x - w, y - w),
(x + w, y + w), (51, 255, 153), 1)
cv2.line(right_E, (x - w, y + w), (x + w, y - w),
(51, 255, 253), 1)
left_E = cv2.line(eye_left_detect, (x - w, y - w),
(x + w, y + w), (51, 255, 153), 1)
cv2.line(left_E, (x - w, y + w), (x + w, y - w),
(51, 255, 253), 1)
face_detected[
eye_coordinates_detect[1][1]:eye_coordinates_detect[1][3],
eye_coordinates_detect[1][0]:eye_coordinates_detect[1]
[2]] = right_E
face_detected[
eye_coordinates_detect[0][1]:eye_coordinates_detect[0][3],
eye_coordinates_detect[0][0]:eye_coordinates_detect[0]
[2]] = left_E
cv2.imshow("Result of the App", cv2.resize(result_app, (600, 600)))
if frame_counter % 5 == 0:
cursor.move(coordi_update_pointer[0], coordi_update_pointer[1])
if key == 27:
break
total_time = time.time() - inference_time_start
total_time_for_inference = round(total_time, 1)
fps = frame_counter / total_time_for_inference
with open(out_path + 'stats.txt', 'w') as f:
f.write('Inference time: ' + str(total_time_for_inference) + '\n')
f.write('FPS: ' + str(fps) + '\n')
f.write('Model load time: ' + str(total_load_time) + '\n')
loger.error("The video stream is over...")
cv2.destroyAllWindows()
feeder_in.close()
def main():
# Grab command line args
args = build_argparser().parse_args()
flags = args.models_outputs_flags
logger = logging.getLogger()
input_file_path = args.input
input_feeder = None
if input_file_path.lower() == "cam":
input_feeder = InputFeeder("cam")
else:
if not os.path.isfile(input_file_path):
logger.error("Unable to find specified video file")
exit(1)
input_feeder = InputFeeder("video", input_file_path)
model_path_dict = {
'FaceDetection': args.face_detection_model,
'FacialLandmarks': args.facial_landmarks_model,
'GazeEstimation': args.gaze_estimation_model,
'HeadPoseEstimation': args.head_pose_estimation_model
}
for file_name_key in model_path_dict.keys():
if not os.path.isfile(model_path_dict[file_name_key]):
logger.error("Unable to find specified " + file_name_key +
" xml file")
exit(1)
fdm = FaceDetection(model_path_dict['FaceDetection'], args.device,
args.cpu_extension)
flm = FacialLandmarks(model_path_dict['FacialLandmarks'], args.device,
args.cpu_extension)
gem = GazeEstimation(model_path_dict['GazeEstimation'], args.device,
args.cpu_extension)
hpem = HeadPoseEstimation(model_path_dict['HeadPoseEstimation'],
args.device, args.cpu_extension)
mc = MouseController('medium', 'fast')
input_feeder.load_data()
fdm.load_model()
flm.load_model()
hpem.load_model()
gem.load_model()
frame_count = 0
for ret, frame in input_feeder.next_batch():
if not ret:
break
frame_count += 1
if frame_count % 5 == 0:
cv2.imshow('video', cv2.resize(frame, (500, 500)))
key = cv2.waitKey(60)
cropped_face, face_coords = fdm.predict(frame, args.prob_threshold)
if type(cropped_face) == int:
logger.error("Unable to detect any face.")
if key == 27:
break
continue
hp_output = hpem.predict(cropped_face)
left_eye_img, right_eye_img, eye_coords = flm.predict(cropped_face)
new_mouse_coord, gaze_vector = gem.predict(left_eye_img, right_eye_img,
hp_output)
if (not len(flags) == 0):
preview_frame = frame
if 'fd' in flags:
preview_frame = cropped_face
if 'fld' in flags:
cv2.rectangle(cropped_face,
(eye_coords[0][0] - 10, eye_coords[0][1] - 10),
(eye_coords[0][2] + 10, eye_coords[0][3] + 10),
(0, 255, 0), 3)
cv2.rectangle(cropped_face,
(eye_coords[1][0] - 10, eye_coords[1][1] - 10),
(eye_coords[1][2] + 10, eye_coords[1][3] + 10),
(0, 255, 0), 3)
if 'hp' in flags:
cv2.putText(
preview_frame,
"Pose Angles: yaw:{:.2f} | pitch:{:.2f} | roll:{:.2f}".
format(hp_output[0], hp_output[1], hp_output[2]), (10, 20),
cv2.FONT_HERSHEY_COMPLEX, 0.25, (0, 255, 0), 1)
if 'ge' in flags:
x, y, w = int(gaze_vector[0] * 12), int(gaze_vector[1] *
12), 160
left_eye = cv2.line(left_eye_img, (x - w, y - w),
(x + w, y + w), (255, 0, 255), 2)
cv2.line(left_eye, (x - w, y + w), (x + w, y - w),
(255, 0, 255), 2)
right_eye = cv2.line(right_eye_img, (x - w, y - w),
(x + w, y + w), (255, 0, 255), 2)
cv2.line(right_eye, (x - w, y + w), (x + w, y - w),
(255, 0, 255), 2)
cropped_face[eye_coords[0][1]:eye_coords[0][3],
eye_coords[0][0]:eye_coords[0][2]] = left_eye
cropped_face[eye_coords[1][1]:eye_coords[1][3],
eye_coords[1][0]:eye_coords[1][2]] = right_eye
cv2.imshow("Visualization", cv2.resize(preview_frame, (500, 500)))
if frame_count % 5 == 0:
mc.move(new_mouse_coord[0], new_mouse_coord[1])
if key == 27:
break
logger.error("VideoStream ended...")
cv2.destroyAllWindows()
input_feeder.close()
def application(args, facedetector, facelm, headpose, gaze):
pointer_controller = MouseController(precision='high', speed='fast')
if args.input != 'CAM':
try:
in_stream = cv2.VideoCapture(args.input)
l = int(in_stream.get(cv2.CAP_PROP_FRAME_COUNT))
webcam = False
if l > 1:
single_image_mode = False
else:
single_image_mode = True
except:
print('Not supported image or video file format. Please pass a supported one.')
exit()
else:
in_stream = cv2.VideoCapture(0)
single_img_mode = False
webcam = True
if not single_image_mode:
count = 0
while(in_stream.isOpened()):
flag, frame = in_stream.read()
if not flag:
break
if count % 25 == 0:
if cv2.waitKey(1) & 0xFF == ord('q'):
break
face_crop, detection = facedetector.get_face_crop(frame, args)
right_eye, left_eye = facelm.get_eyes_coordinates(face_crop)
right_eye_crop, left_eye_crop, right_eye_coords, left_eye_coords = helpers.get_eyes_crops(face_crop, right_eye,left_eye)
headpose_angles = headpose.get_headpose_angles(face_crop)
(x_movement, y_movement), gaze_vector = gaze.get_gaze(right_eye_crop, left_eye_crop, headpose_angles)
if args.show_face:
frame = cv2.rectangle(frame,
(detection[0],detection[1]),
(detection[2],detection[3]),
color=(0,255,0),
thickness=5)
if args.show_headpose:
frame = cv2.putText(frame, 'Roll: '+
str(headpose_angles[2])+' '+
'Pitch: '+str(headpose_angles[1])+' '+
'Yaw: '+str(headpose_angles[0]),(15,20),cv2.FONT_HERSHEY_SIMPLEX,0.65,(0,0,0),2)
if args.show_eyes:
right_eye_coords = [right_eye_coords[0]+detection[1], right_eye_coords[1]+detection[1],
right_eye_coords[2]+detection[0], right_eye_coords[3]+detection[0]]
left_eye_coords = [left_eye_coords[0]+detection[1], left_eye_coords[1]+detection[1],
left_eye_coords[2]+detection[0], left_eye_coords[3]+detection[0]]
frame = cv2.rectangle(frame,
(right_eye_coords[2],right_eye_coords[1]),
(right_eye_coords[3],right_eye_coords[0]),
color=(255,0,0),
thickness=5)
frame = cv2.rectangle(frame,
(left_eye_coords[2],left_eye_coords[1]),
(left_eye_coords[3],left_eye_coords[0]),
color=(255,0,0),
thickness=5)
if args.show_gaze:
# Right eye:
x_r_eye = int(right_eye[0]*face_crop.shape[1]+detection[0])
y_r_eye = int(right_eye[1]*face_crop.shape[0]+detection[1])
x_r_shift, y_r_shift = int(x_r_eye+gaze_vector[0]*100), int(y_r_eye-gaze_vector[1]*100)
# Left eye:
x_l_eye = int(left_eye[0]*face_crop.shape[1]+detection[0])
y_l_eye = int(left_eye[1]*face_crop.shape[0]+detection[1])
x_l_shift, y_l_shift = int(x_l_eye+gaze_vector[0]*100), int(y_l_eye-gaze_vector[1]*100)
frame = cv2.arrowedLine(frame, (x_r_eye, y_r_eye), (x_r_shift, y_r_shift), (0, 255, 0), 4)
frame = cv2.arrowedLine(frame, (x_l_eye, y_l_eye), (x_l_shift, y_l_shift), (0, 255, 0), 4)
cv2.namedWindow('Output',cv2.WINDOW_NORMAL)
cv2.resizeWindow('Output', 800,600)
cv2.imshow('Output', frame)
#pointer_controller.move(x_movement,y_movement)
count = count + 1
in_stream.release()
cv2.destroyAllWindows()
def main(args):
# getting the arguments
if args.get_perf_counts.lower() == "true":
perf_counts = True
elif args.get_perf_counts.lower() == "false":
perf_counts = False
precision = args.precision.lower()
speed = args.speed.lower()
media_type = args.media_type.lower()
media_path = args.media_file
toggle_ui = args.show_video
print(toggle_ui)
batch_size = args.batch_size
device = args.device
iterations = 1 if media_type == "cam" else int(args.iterations)
#initialize the mouse object
mouse = MouseController(precision, speed)
# Initialize the input feeder
feed = InputFeeder(media_type, batch_size, media_path)
# Initialize and load the inference models
model = Model(face_detection, facial_landmarks, gaze_estimation,
head_pose_estimation, device)
model.load_models()
for _ in range(iterations):
feed.load_data()
#This will be used as a way to keep track of the average time for the preprocessing and inference of the models
times = np.zeros((8, ))
counter_frames = 0
if media_type != "image":
width = feed.cap.get(3)
height = feed.cap.get(4)
else:
height, width, _ = feed.cap.shape
try:
for frame in feed.next_batch(media_type):
counter_frames += 1
#generates the prediction
x, y, gaze_vector, times = model.predict(
frame, width, height, times)
#generates the movement on the cursor
mouse.move(x, y)
if perf_counts:
cv2.putText(
frame, "Preprocess Face Detection: " +
str(times[0] / counter_frames * 1000) + " ms", (0, 50),
cv2.FONT_HERSHEY_SIMPLEX, 1, (209, 80, 0), 3)
cv2.putText(
frame, "Inference Face Detection: " +
str(times[1] / counter_frames * 1000) + " ms",
(0, 100), cv2.FONT_HERSHEY_SIMPLEX, 1, (209, 80, 0), 3)
cv2.putText(
frame, "Preprocess Facial Landmarks: " +
str(times[2] / counter_frames * 1000) + " ms",
(0, 150), cv2.FONT_HERSHEY_SIMPLEX, 1, (209, 80, 0), 3)
cv2.putText(
frame, "Inference Facial Landmarks: " +
str(times[3] / counter_frames * 1000) + " ms",
(0, 200), cv2.FONT_HERSHEY_SIMPLEX, 1, (209, 80, 0), 3)
cv2.putText(
frame, "Preprocess Head Pose: " +
str(times[4] / counter_frames * 1000) + " ms",
(0, 250), cv2.FONT_HERSHEY_SIMPLEX, 1, (209, 80, 0), 3)
cv2.putText(
frame, "Inference Head Pose: " +
str(times[5] / counter_frames * 1000) + " ms",
(0, 300), cv2.FONT_HERSHEY_SIMPLEX, 1, (209, 80, 0), 3)
cv2.putText(
frame, "Preprocess Gaze Estimation: " +
str(times[6] / counter_frames * 1000) + " ms",
(0, 350), cv2.FONT_HERSHEY_SIMPLEX, 1, (209, 80, 0), 3)
cv2.putText(
frame, "Inference Gaze Estimation: " +
str(times[7] / counter_frames * 1000) + " ms",
(0, 400), cv2.FONT_HERSHEY_SIMPLEX, 1, (209, 80, 0), 3)
print("Preprocess Face Detection: " +
str(times[0] / counter_frames * 1000) + " ms")
print("Inference Face Detection: " +
str(times[1] / counter_frames * 1000) + " ms")
print("Preprocess Facial Landmarks: " +
str(times[2] / counter_frames * 1000) + " ms")
print("Inference Facial Landmarks: " +
str(times[3] / counter_frames * 1000) + " ms")
print("Preprocess Head Pose: " +
str(times[4] / counter_frames * 1000) + " ms")
print("Inference Head Pose: " +
str(times[5] / counter_frames * 1000) + " ms")
print("Preprocess Gaze Estimation: " +
str(times[6] / counter_frames * 1000) + " ms")
print("Inference Gaze Estimation: " +
str(times[7] / counter_frames * 1000) + " ms")
if toggle_ui == True:
cv2.imshow("Frame", frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if cv2.waitKey(1) & 0xFF == ord('i'):
toggle_UI = False if toggle_UI else True
except:
print("Video has ended or couldn't continue")
if perf_counts:
print("Final average: ")
print("Preprocess Face Detection: " +
str(times[0] / counter_frames * 1000) + " ms")
print("Inference Face Detection: " +
str(times[1] / counter_frames * 1000) + " ms")
print("Preprocess Facial Landmarks: " +
str(times[2] / counter_frames * 1000) + " ms")
print("Inference Facial Landmarks: " +
str(times[3] / counter_frames * 1000) + " ms")
print("Preprocess Head Pose: " +
str(times[4] / counter_frames * 1000) + " ms")
print("Inference Head Pose: " +
str(times[5] / counter_frames * 1000) + " ms")
print("Preprocess Gaze Estimation: " +
str(times[6] / counter_frames * 1000) + " ms")
print("Inference Gaze Estimation: " +
str(times[7] / counter_frames * 1000) + " ms")
feed.close()
cv2.destroyAllWindows()
def infer(self, args):
# Create instances from the models' classes
FDM_net = ModelFaceDetection()
HPE_net = ModelHeadPoseEstimation()
FLD_net = ModelFacialLandmarksDetection()
GEM_net = ModelGazeEstimation()
mouse_controller = MouseController('high', 'fast')
# Load the models
start1 = time.time()
FDM_net.load_model(args.face_detection_model, args.device)
FDM_load_t = time.time() - start1
start2 = time.time()
HPE_net.load_model(args.head_pose_estimation_model, args.device)
HPE_load_t = time.time() - start2
start3 = time.time()
FLD_net.load_model(args.facial_landmarks_detection_model, args.device)
FLD_load_t = time.time() - start3
start4 = time.time()
GEM_net.load_model(args.gaze_estimation_model, args.device)
GEM_load_t = time.time() - start4
print('All models are loaded!')
#Check the inputs
# To make the mouse moving we need video stream either from camera or video path
if args.input.lower() == 'cam':
# Initialise the InputFeeder class
input_feeder = InputFeeder(input_type='cam', input_file=args.input)
else:
if not os.path.isfile(args.input):
log.error("Please insert valid video path to run the app.")
exit()
# Initialise the InputFeeder class
input_feeder = InputFeeder(input_type='video',
input_file=args.input)
# Load the video capture
input_feeder.load_data()
# Inference time
inference = time.time()
# Read from the video capture
for flag, frame in input_feeder.next_batch():
if not flag:
break
key_pressed = cv2.waitKey(60)
# Run inference on the models
start5 = time.time()
face_coords = FDM_net.predict(frame)
FDM_infer_t = time.time() - start5
# crop the face from the frame
cropped_face = frame[face_coords[1]:face_coords[3],
face_coords[0]:face_coords[2]]
#Everything depends on the face detection output, if no face detected then repeat
if len(face_coords) == 0:
log.error("There is no faces detected.")
continue
start6 = time.time()
HP_angles = HPE_net.predict(cropped_face, face_coords)
HPE_infer_t = time.time() - start6
if args.display_flag:
#### display the face
O_frame = cv2.rectangle(frame.copy(),
(face_coords[0], face_coords[1]),
(face_coords[2], face_coords[3]),
(255, 255, 0), 2)
#### display the pose angles
# Link for pose estimation output code resource: https://sudonull.com/post/6484-Intel-OpenVINO-on-Raspberry-Pi-2018-harvest
cos_r = cos(HP_angles[2] * pi / 180)
sin_r = sin(HP_angles[2] * pi / 180)
cos_y = cos(HP_angles[0] * pi / 180)
sin_y = sin(HP_angles[0] * pi / 180)
cos_p = cos(HP_angles[1] * pi / 180)
sin_p = sin(HP_angles[1] * pi / 180)
x = int((face_coords[0] + face_coords[2]) / 2)
y = int((face_coords[1] + face_coords[3]) / 2)
cv2.line(O_frame, (x, y),
(x + int(65 *
(cos_r * cos_y + sin_y * sin_p * sin_r)),
y + int(65 * cos_p * sin_r)), (255, 0, 0),
thickness=2)
cv2.line(O_frame, (x, y),
(x + int(65 *
(cos_r * sin_y * sin_p + cos_y * sin_r)),
y - int(65 * cos_p * cos_r)), (0, 255, 0),
thickness=2)
cv2.line(O_frame, (x, y),
(x + int(65 * sin_y * cos_p), y + int(65 * sin_p)),
(0, 0, 255),
thickness=2)
start7 = time.time()
l_e, r_e, l_e_image, r_e_image, e_center = FLD_net.predict(
O_frame, cropped_face, face_coords)
FLD_infer_t = time.time() - start7
###display landmarks for both eyes
if args.display_flag:
cv2.circle(O_frame,
(face_coords[0] + l_e[0], face_coords[1] + l_e[1]),
29, (0, 255, 255), 2)
cv2.circle(O_frame,
(face_coords[0] + r_e[0], face_coords[1] + r_e[1]),
29, (0, 255, 255), 2)
start8 = time.time()
g_vec = GEM_net.predict(l_e_image, r_e_image, HP_angles)
GEM_infer_t = time.time() - start8
###display gaze model output
if args.display_flag:
cv2.arrowedLine(O_frame,
(int(e_center[0][0]), int(e_center[0][1])),
(int(e_center[0][0]) + int(g_vec[0] * 90),
int(e_center[0][1]) + int(-g_vec[1] * 90)),
(203, 192, 255), 2)
cv2.arrowedLine(O_frame,
(int(e_center[1][0]), int(e_center[1][1])),
(int(e_center[1][0]) + int(g_vec[0] * 90),
int(e_center[1][1]) + int(-g_vec[1] * 90)),
(203, 192, 255), 2)
# change the pointer position according to the estimated gaze direction
mouse_controller.move(g_vec[0], g_vec[1])
if key_pressed == 27:
break
# Display the resulting frame
cv2.imshow('Mouse Controller App Results',
cv2.resize(O_frame, (750, 550)))
inference_time = time.time() - inference
print("Loading time: \n1-Face detection: " + str(FDM_load_t) +
"\n2- Head pose estimation: " + str(HPE_load_t) +
"\n3-Facial landmarks model: " + str(FLD_load_t) +
"\n4-Gaze estimation model: " + str(GEM_load_t))
print("Output inference time: \n1-Face detection: " +
str(FDM_infer_t) + "\n2- Head pose estimation: " +
str(HPE_infer_t) + "\n3-Facial landmarks model: " +
str(FLD_infer_t) + "\n4-Gaze estimation model: " +
str(GEM_infer_t))
# close the input feeder and destroy all opened windows
input_feeder.close()
cv2.destroyAllWindows
def main():
args = build_argparser().parse_args()
frame_num = 0
inference_time = 0
counter = 0
# Initialize the Inference Engine
fd = FaceDetection()
fld = Facial_Landmarks_Detection()
ge = Gaze_Estimation()
hp = Head_Pose_Estimation()
# Load Models
fd.load_model(args.face_detection_model, args.device, args.cpu_extension)
fld.load_model(args.facial_landmark_model, args.device, args.cpu_extension)
ge.load_model(args.gaze_estimation_model, args.device, args.cpu_extension)
hp.load_model(args.head_pose_model, args.device, args.cpu_extension)
# Mouse Controller precision and speed
mc = MouseController('medium', 'fast')
# feed input from an image, webcam, or video to model
if args.input == "cam":
feed = InputFeeder("cam")
else:
assert os.path.isfile(args.input), "Specified input file doesn't exist"
feed = InputFeeder("video", args.input)
feed.load_data()
frame_count = 0
for frame in feed.next_batch():
frame_count += 1
inf_start = time.time()
if frame is not None:
try:
key = cv2.waitKey(60)
det_time = time.time() - inf_start
# make predictions
detected_face, face_coords = fd.predict(
frame.copy(), args.prob_threshold)
hp_output = hp.predict(detected_face.copy())
left_eye, right_eye, eye_coords = fld.predict(
detected_face.copy())
new_mouse_coord, gaze_vector = ge.predict(
left_eye, right_eye, hp_output)
stop_inference = time.time()
inference_time = inference_time + stop_inference - inf_start
counter = counter + 1
# Visualization
preview = args.visualization
if preview:
preview_frame = frame.copy()
face_frame = detected_face.copy()
draw_face_bbox(preview_frame, face_coords)
display_hp(preview_frame, hp_output, face_coords)
draw_landmarks(face_frame, eye_coords)
draw_gaze(face_frame, gaze_vector, left_eye.copy(),
right_eye.copy(), eye_coords)
if preview:
img = np.hstack((cv2.resize(preview_frame, (500, 500)),
cv2.resize(face_frame, (500, 500))))
else:
img = cv2.resize(frame, (500, 500))
cv2.imshow('Visualization', img)
# set speed
if frame_count % 5 == 0:
mc.move(new_mouse_coord[0], new_mouse_coord[1])
# INFO
log.info("NUMBER OF FRAMES: {} ".format(frame_num))
log.info("INFERENCE TIME: {}ms".format(det_time * 1000))
frame_num += 1
if key == 27:
break
except:
print(
'Not supported image or video file format. Please send in a supported video format.'
)
exit()
feed.close()
def main():
args = build_argparser().parse_args()
logger = logging.getLogger('main')
model_path_dict = {
'FaceDetectionModel': args.faceDetectionModel,
'FacialLandmarksModel': args.facialLandmarksModel,
'HeadPoseEstimationModel': args.headPoseEstimationModel,
'GazeEstimationModel': args.gazeEstimationModel
}
bbox_flag = args.bbox_flag
input_filename = args.input
device_name = args.device
prob_threshold = args.prob_threshold
output_path = args.output_path
if input_filename.lower() == 'cam':
feeder = InputFeeder(input_type='cam')
else:
if not os.path.isfile(input_filename):
logger.error("Unable to find specified video file")
exit(1)
feeder = InputFeeder(input_type='video', input_file=input_filename)
for model_path in list(model_path_dict.values()):
if not os.path.isfile(model_path):
logger.error("Unable to find specified model file" +
str(model_path))
exit(1)
face_detection_model = Face_detection(
model_path_dict['FaceDetectionModel'],
device_name,
threshold=prob_threshold)
facial_landmarks_detection_model = Landmark_Detection(
model_path_dict['FacialLandmarksModel'],
device_name,
threshold=prob_threshold)
head_pose_estimation_model = Head_pose(
model_path_dict['HeadPoseEstimationModel'],
device_name,
threshold=prob_threshold)
gaze_estimation_model = Gaze_estimation(
model_path_dict['GazeEstimationModel'],
device_name,
threshold=prob_threshold)
is_benchmarking = False
if not is_benchmarking:
mouse_controller = MouseController('medium', 'fast')
start_model_load_time = time.time()
face_detection_model.load_model()
facial_landmarks_detection_model.load_model()
head_pose_estimation_model.load_model()
gaze_estimation_model.load_model()
total_model_load_time = time.time() - start_model_load_time
feeder.load_data()
out_video = cv2.VideoWriter(os.path.join('output_video.mp4'),
cv2.VideoWriter_fourcc(*'avc1'),
int(feeder.get_fps() / 10), (1920, 1080), True)
frame_count = 0
start_inference_time = time.time()
for ret, frame in feeder.next_batch():
if not ret:
break
frame_count += 1
key = cv2.waitKey(60)
try:
face_coords, image_copy = face_detection_model.predict(frame)
if type(image_copy) == int:
logger.warning("Unable to detect the face")
if key == 27:
break
continue
left_eye, right_eye, eye_coords = facial_landmarks_detection_model.predict(
image_copy)
hp_output = head_pose_estimation_model.predict(image_copy)
mouse_coords, gaze_coords = gaze_estimation_model.predict(
left_eye, right_eye, hp_output)
except Exception as e:
logger.warning("Could predict using model" + str(e) +
" for frame " + str(frame_count))
continue
image = cv2.resize(frame, (500, 500))
if not len(bbox_flag) == 0:
bbox_frame = draw_bbox(frame, bbox_flag, image_copy, left_eye,
right_eye, face_coords, eye_coords,
hp_output, gaze_coords)
image = np.hstack(
(cv2.resize(frame,
(500, 500)), cv2.resize(bbox_frame, (500, 500))))
cv2.imshow('preview', image)
out_video.write(frame)
if frame_count % 5 == 0 and not is_benchmarking:
mouse_controller.move(mouse_coords[0], mouse_coords[1])
if key == 27:
break
total_time = time.time() - start_inference_time
total_inference_time = round(total_time, 1)
fps = frame_count / total_inference_time
try:
os.mkdir(output_path)
except OSError as error:
logger.error(error)
with open(output_path + 'stats.txt', 'w') as f:
f.write(str(total_inference_time) + '\n')
f.write(str(fps) + '\n')
f.write(str(total_model_load_time) + '\n')
logger.info('Model load time: ' + str(total_model_load_time))
logger.info('Inference time: ' + str(total_inference_time))
logger.info('FPS: ' + str(fps))
logger.info('Video stream ended')
cv2.destroyAllWindows()
feeder.close()
class MoveMouse:
'''
Main Class for the Mouse Controller app.
This is the class where all the models are stitched together to control the mouse pointer
'''
def __init__(self, args):
'''
This method instances variables for the Facial Landmarks Detection Model.
Args:
args = All arguments parsed by the arguments parser function
Return:
None
'''
init_start_time = time.time()
self.output_path = args.output_path
self.show_output = args.show_output
self.total_processing_time = 0
self.count_batch = 0
self.inference_speed = []
self.avg_inference_speed = 0
if args.all_devices != 'CPU':
args.face_device = args.all_devices
args.face_landmark_device = args.all_devices
args.head_pose_device = args.all_devices
args.gaze_device = args.all_devices
model_init_start = time.time()
self.face_model = FaceDetection(args.face_model, args.face_device,
args.face_device_ext,
args.face_prob_threshold)
self.landmarks_model = FacialLandmarksDetection(
args.face_landmark_model, args.face_landmark_device,
args.face_landmark_device_ext, args.face_landmark_prob_threshold)
self.head_pose_model = HeadPoseEstimation(
args.head_pose_model, args.head_pose_device,
args.head_pose_device_ext, args.head_pose_prob_threshold)
self.gaze_model = GazeEstimation(args.gaze_model, args.gaze_device,
args.gaze_device_ext,
args.gaze_prob_threshold)
self.model_init_time = time.time() - model_init_start
log.info('[ Main ] All required models initiallized')
self.mouse_control = MouseController(args.precision, args.speed)
log.info('[ Main ] Mouse controller successfully initialized')
self.input_feeder = InputFeeder(args.batch_size, args.input_type,
args.input_file)
log.info('[ Main ] Initialized input feeder')
model_load_start = time.time()
self.face_model.load_model()
self.landmarks_model.load_model()
self.head_pose_model.load_model()
self.gaze_model.load_model()
self.model_load_time = time.time() - model_load_start
self.app_init_time = time.time() - init_start_time
log.info('[ Main ] All moadels loaded to Inference Engine\n')
return None
def draw_face_box(self, frame, face_coords):
'''
Draws face's bounding box on the input frame
Args:
frame = Input frame from video or camera feed. It could also be an input image
Return:
frame = Frame with bounding box of faces drawn on it
'''
start_point = (face_coords[0][0], face_coords[0][1])
end_point = (face_coords[0][2], face_coords[0][3])
thickness = 5
color = (255, 86, 0)
frame = cv2.rectangle(frame, start_point, end_point, color, thickness)
return frame
def draw_eyes_boxes(self, frame, left_eye_coords, right_eye_coords):
'''
Draws face's bounding box on the input frame
Args:
frame = Input frame from video or camera feed. It could also be an input image
Return:
frame = Frame with bounding box of left and right eyes drawn on it
'''
left_eye_start_point = (left_eye_coords[0], left_eye_coords[1])
left_eye_end_point = (left_eye_coords[2], left_eye_coords[3])
right_eye_start_point = (right_eye_coords[0], right_eye_coords[1])
right_eye_end_point = (right_eye_coords[2], right_eye_coords[3])
thickness = 5
color = (0, 210, 0)
frame = cv2.rectangle(frame, left_eye_start_point, left_eye_end_point,
color, thickness)
frame = cv2.rectangle(frame, right_eye_start_point,
right_eye_end_point, color, thickness)
return frame
def draw_outputs(self, frame):
'''
Draws the inference outputs (bounding boxes of the face and both eyes and
the 3D head pose directions) of the four models onto the frames.
Args:
frame = Input frame from video or camera feed. It could also be an input image
Return:
frame = Frame with all inference outputs drawn on it
'''
frame = self.draw_face_box(frame, self.face_coords)
frame = self.draw_eyes_boxes(frame, self.left_eye_coords,
self.right_eye_coords)
frame_id = f'Batch id = {self.count_batch}'
avg_inference_speed = f'Avg. inference speed = {self.avg_inference_speed:.3f}fps'
total_processing_time = f'Total infer. time = {self.total_processing_time:.3f}s'
cv2.putText(frame, frame_id, (15, 15), cv2.FONT_HERSHEY_COMPLEX, 0.45,
(255, 86, 0), 1)
cv2.putText(frame, avg_inference_speed, (15, 30),
cv2.FONT_HERSHEY_COMPLEX, 0.45, (255, 86, 0), 1)
cv2.putText(frame, total_processing_time, (15, 45),
cv2.FONT_HERSHEY_COMPLEX, 0.45, (255, 86, 0), 1)
return frame
def run_inference(self, frame):
'''
Performs inference on the input video or image by passing it through all four
models to get the desired coordinates for moving the mouse pointer.
Args:
frame = Input image, frame from video or camera feed
Return:
None
'''
self.input_feeder.load_data()
for frame in self.input_feeder.next_batch():
if self.input_feeder.frame_flag == True:
log.info('[ Main ] Started processing a new batch')
start_inference = time.time()
self.face_coords, self.face_crop = self.face_model.predict(
frame)
if self.face_coords == []:
log.info(
'[ Main ] No face detected.. Waiting for you to stare at the camera'
)
f.write('[ Error ] No face was detected')
else:
self.head_pose_angles = self.head_pose_model.predict(
self.face_crop)
self.left_eye_coords, self.left_eye_image, self.right_eye_coords, self.right_eye_image = self.landmarks_model.predict(
self.face_crop)
self.x, self.y = self.gaze_model.predict(
self.left_eye_image, self.right_eye_image,
self.head_pose_angles)
log.info(
f'[ Main ] Relative pointer coordinates: [{self.x:.2f}, {self.y:.2f}]'
)
batch_process_time = time.time() - start_inference
self.total_processing_time += batch_process_time
self.count_batch += 1
log.info(
f'[ Main ] Finished processing batch. Time taken = {batch_process_time}s\n'
)
self.mouse_control.move(self.x, self.y)
if self.show_output:
self.draw_outputs(frame)
cv2.imshow('Computer Pointer Controller Output', frame)
self.inference_speed.append(self.count_batch /
self.total_processing_time)
self.avg_inference_speed = sum(self.inference_speed) / len(
self.inference_speed)
with open(os.path.join(self.output_path, 'outputs.txt'),
'w+') as f:
f.write('INFERENCE STATS\n')
f.write(
f'Total model initialization time : {self.model_init_time:.2f}s\n'
)
f.write(
f'Total model load time: {self.model_load_time:.2f}s\n'
)
f.write(
f'App initialization time: {self.app_init_time:.2f}s\n'
)
f.write(
f'Total processing time: {self.total_processing_time:.2f}s\n'
)
f.write(
f'Average inference speed: {self.avg_inference_speed:.2f}FPS\n'
)
f.write(f'Batch count: {self.count_batch}\n\n')
f.write('LAST OUTPUTS\n')
f.write(f'Face coordinates: {self.face_coords}\n')
f.write(f'Left eye coordinates: {self.left_eye_coords}\n')
f.write(
f'Right eye coordinates: {self.right_eye_coords}\n')
f.write(f'Head pose angles: {self.head_pose_angles}\n')
f.write(
f'Relative pointer coordinates/ Gaze vector: [{self.x:.2f}, {self.y:.2f}]'
)
else:
self.input_feeder.close()
cv2.destroyAllWindows()
log.info(
f'[ Main ] All input Batches processed in {self.total_processing_time:.2f}s'
)
log.info('[ Main ] Shutting down app...')
log.info('[ Main ] Mouse controller app has been shut down.')
break
return
