def process_video(file_input, file_output, display_intermediate_output):
if file_input is None:
feed = InputFeeder(input_type='cam')
else:
feed = InputFeeder(input_type='video', input_file=file_input)
feed.load_data()
w = int(feed.cap.get(cv2.CAP_PROP_FRAME_WIDTH))
h = int(feed.cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = int(feed.cap.get(cv2.CAP_PROP_FPS))
out = cv2.VideoWriter(file_output, cv2.VideoWriter_fourcc(*'avc1'), fps, (w, h), True)
frame_counter = 0
for batch in feed.next_batch():
frame_counter += 1
result, frame = process_single_frame(batch, display_intermediate_output)
out.write(frame)
logging.debug(f'Frame #{frame_counter} result: {result}')
if type(result) == str and result == 'No face detected':
logging.warning('Frame {}: No face detected', frame_counter)
if mouse_controller is not None:
mouse_controller.move(result[0], result[1])
out.release()
feed.close()
def main():
args = build_argparser().parse_args()
inputFile = args.input
inputFeeder = None
if inputFile.lower() == "cam":
inputFeeder = InputFeeder("cam")
else:
if not os.path.isfile(inputFile):
print("Unable to find input file")
exit(1)
inputFeeder = InputFeeder("video",inputFile)
start_model_loading = time.time()
detect,landmark,gaze,pose=init_models(args)
inputFeeder.load_data()
LoadModel(detect, landmark, gaze, pose)
model_loading_time = time.time() - start_model_loading
frame_count,inference_time = inference_frame(detect,pose,landmark,gaze,inputFeeder,args)
fps = frame_count / inference_time
print("video is complete!")
print(f'Model took {model_loading_time} s to load')
print(f'Inference time of the model is: {inference_time} s')
print(f'Average inference time of the model is : {inference_time/frame_count} s')
print(f'FPS is {fps/5} frame/second')
cv2.destroyAllWindows()
inputFeeder.close()
def cam_or_video(inputs, inputFeeder):
if inputs.lower() == "cam":
inputFeeder = InputFeeder("cam")
if not os.path.isfile(inputs):
print("Unable to find input file")
exit(1)
inputFeeder = InputFeeder("video", inputs)
return inputFeeder
def init_feeder(args):
input_feeder = None
if args.input.lower() == "cam":
input_feeder = InputFeeder("cam")
else:
if not os.path.isfile(args.input):
logging.error("Unable to find specified video file")
exit(1)
input_feeder = InputFeeder("video", args.input)
return input_feeder
def input_feeder_func(input_file_path):
# Checks for live feed
if input_file_path == 'CAM':
input_feeder = InputFeeder("cam")
# Checks for video file
else:
input_feeder = InputFeeder("video", input_file_path)
assert os.path.isfile(
input_file_path), "Specified input file doesn't exist"
return input_feeder
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 check_source(filepath, logger):
feeder = None
if filepath.lower() == 'cam':
feeder = InputFeeder('cam')
else:
if not os.path.isfile(filepath):
logger.error("Unable to find specified video file")
exit(1)
feeder = InputFeeder('video', filepath)
return feeder
def feedInput(input):
feeder = None
input_type = "cam"
if input != 'CAM':
assert os.path.isfile(input)
if input.endswith(('.jpg', '.bmp', '.png')):
input_type = "image"
else:
input_type = "video"
feeder = InputFeeder(input_type=input_type, input_file=input)
else:
feeder = InputFeeder(input_type=input_type)
return feeder
def run_inference(args):
feed = InputFeeder(input_type='video', input_file=args.input)
feed.load_data()
for batch in feed.next_batch():
cv2.imshow("Output", cv2.resize(batch, (500, 500)))
key = cv2.waitKey(60)
if (key == 27):
break
# getting face
faceDetection = FaceDetection(model_name=args.face_detection_model)
faceDetection.load_model()
face = faceDetection.predict(batch)
# getting eyes
facialLandmarksDetection = FacialLandmarksDetection(
args.facial_landmarks_detection_model)
facialLandmarksDetection.load_model()
left_eye, right_eye = facialLandmarksDetection.predict(face)
# getting head pose angles
headPoseEstimation = HeadPoseEstimation(
args.head_pose_estimation_model)
headPoseEstimation.load_model()
head_pose = headPoseEstimation.predict(face)
print("head pose angles: ", head_pose)
# get mouse points
gazeEstimation = GazeEstimation(args.gaze_estimation_model)
gazeEstimation.load_model()
mouse_coords = gazeEstimation.predict(left_eye, right_eye, head_pose)
print("gaze output: ", mouse_coords)
feed.close()
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 process_video(input_video, video_output, visualize):
if input_video is None:
feed = InputFeeder(input_type='cam')
else:
feed = InputFeeder(input_type='video', input_file=input_video)
feed.load_data()
w = int(feed.cap.get(cv2.CAP_PROP_FRAME_WIDTH))
h = int(feed.cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = int(feed.cap.get(cv2.CAP_PROP_FPS))
fps = int(fps / 4)
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter(video_output, fourcc, fps, (w, h), True)
frame_counter = 0
for frame in feed.next_batch():
if frame is not None:
frame_counter += 1
key = cv2.waitKey(10)
result, output_frame = process_frame(frame, visualize)
out.write(output_frame)
print("Frame: {} result: {}".format(frame_counter, result))
logger.info("Frame: {} result: {}".format(frame_counter, result))
esc_code = 27
if key == esc_code:
break
if mouse_controller is not None:
try:
mouse_controller.move(result[0], result[1])
except Exception as e:
print("Mouse controller exception:\n", e)
logger.info("Mouse controller exception:{}".format(e))
else:
break
cv2.destroyAllWindows()
out.release()
feed.close()
print("Saved the video")
logger.info("Saved the 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 createInputFeeder(cmd_paras):
if cmd_paras.input_type==CAM_INPUT_TYPE:
input_file = None
else:
input_file = cmd_paras.input_file
input_feeder=InputFeeder(input_type=cmd_paras.input_type, input_file=input_file)
return input_feeder
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 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 process_image(file_path, file_output, display_intermediate_output):
feed = InputFeeder(input_type='image', input_file=file_path)
feed.load_data()
for batch in feed.next_batch():
result, image = process_single_frame(batch, display_intermediate_output)
# cv2.imshow('demo image', image)
cv2.imwrite(file_output, image)
cv2.waitKey(0)
cv2.destroyAllWindows()
feed.close()
def main(args):
inference = Inference(args.model)
inference.load_model()
input = args.input
if input == 0:
input_feeder = InputFeeder('cam', input)
elif input.endswith('.jpg') or input.endswith('.jpeg') or input.endswith(
'.bmp'):
input_feeder = InputFeeder('image', input)
is_image = True
else:
input_feeder = InputFeeder('video', input)
input_feeder.load_data()
if is_image:
outputs = inference.predict(input_feeder.cap)
inference.preprocess_output(outputs)
return 0
frames = 0
for ret, frame in input_feeder.next_batch():
if not ret:
break
frames += 1
key = cv2.waitKey(60)
if key == 27:
break
outputs = inference.predict(frame)
inference.preprocess_output(outputs)
input_feeder.close()
def main():
args = build_argparser().parse_args()
visual_flags = args.flag_visualization
input_path = args.input
Dict_model_path = {
'Face': args.face_detection_path,
'Landmarks': args.facial_landmarks_path,
'Headpose': args.head_pose_path,
'Gaze': args.gaze_estimation_path
}
if input_path == "CAM" or input_path=="cam":
print("\n## You are using CAMERA right now..." + input_path.lower() + " detected!")
input_feeder = InputFeeder(input_path.lower())
else:
if not os.path.isfile(input_path):
print("\n## Input file not exists in Path: " + input_path + ". Please check again !!!")
exit(1)
else:
print('\n## Input path exists: '+ input_path + '\n')
input_feeder = InputFeeder("video", input_path)
for model_key in Dict_model_path.keys():
print(Dict_model_path[model_key])
if not os.path.isfile(Dict_model_path[model_key]):
print("\n## " + model_key + " Model path not exists: " + Dict_model_path[model_key] + ' Please try again !!!')
exit(1)
else:
print('## '+model_key + " Model path is correct: " + Dict_model_path[model_key])
print(input_feeder)
print(input_path)
print(visual_flags)
print(args.cpu_extension)
print(args.device)
print(args.prob_threshold)
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 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 run(self, args):
inputFeeder = InputFeeder(args.input)
i = 0
objectsDetection = ObjectsDetection()
frame = None
while self.execute:
try:
frame = next(inputFeeder.next_batch())
except StopIteration:
logging.error('Failed to obtain input stream.')
break
if frame is None:
break
objectsDetection.inputs(frame)
objectsDetection.wait()
outputs = objectsDetection.outputs()
print(frame)
i = i + 1
print(i)
inputFeeder.close()
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():
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()
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()
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 infer_on_stream(args, model):
'''
:param args: argparser arguments
:param model: loaded model
'''
# get the loaded model instance
objectDetection = model
# Handle the input stream
# Check if the input is a webcam or video or image
if args.input == 'cam':
feed = InputFeeder(input_type='cam', flip=1)
feed.set_camera_properties(args.width, args.height, args.fps)
elif args.input == 'picam':
feed = InputFeeder(input_type='picam')
feed.set_camera_properties(args.width, args.height, args.fps)
elif args.input.endswith('.jpg') or args.input.endswith(
'.bmp') or args.input.endswith('.png'):
feed = InputFeeder(input_type='image', input_file=args.input)
elif args.input.endswith('.mp4'):
feed = InputFeeder(input_type='video', input_file=args.input)
else:
print(
"ERROR: Invalid input, it must be CAM, image (.jpg, .bmp or .png) or video (.mp4)!"
)
raise NotImplementedError
feed.load_data()
# run-time switches
ui_marking = True
fps_marking = False
label_background_color = (125, 175, 75)
label_text_color = (255, 255, 255) # white text
cv2.namedWindow("Frame", cv2.WINDOW_NORMAL)
cv2.setWindowProperty("Frame", cv2.WND_PROP_FULLSCREEN,
cv2.WINDOW_FULLSCREEN)
# Start recording of output saving is enabled
if args.save_output:
now = datetime.datetime.now()
out = cv2.VideoWriter(now.strftime("out-%Y%m%d-%H%M%S.avi"),
cv2.VideoWriter_fourcc(*'MJPG'), 15,
(args.width, args.height))
for batch in feed.next_batch():
if batch is None:
continue
# start measuring overall execution time
start_processing_time = time.time()
# 1) First detect objects on the image
start_object_infer_time = time.time() # time measurement started
objects = objectDetection.predict(batch)
total_object_infer_time = time.time(
) - start_object_infer_time # time measurement finished
# executed only if there are objects on the image
if len(objects) > 0:
# if UI marking is turned on draw the vectors, rectangles, etc
if ui_marking:
# objects bounding boxes
for obj in objects:
# draw the bounding box
cv2.rectangle(batch, (obj['xmin'], obj['ymin']),
(obj['xmax'], obj['ymax']), obj['color'], 2)
# prepare the label
label_text = f"{obj['class']}: {obj['confidence']*100:.3}%"
label_size = cv2.getTextSize(label_text,
cv2.FONT_HERSHEY_SIMPLEX, 0.8,
1)[0]
label_left = obj['xmin']
label_top = obj['ymin'] - label_size[1]
if (label_top < 1):
label_top = 1
label_right = label_left + label_size[0]
label_bottom = label_top + label_size[1] - 3
cv2.rectangle(batch, (label_left - 1, label_top - 6),
(label_right + 1, label_bottom + 1),
label_background_color, -1)
cv2.putText(batch, label_text, (label_left, label_bottom),
cv2.FONT_HERSHEY_SIMPLEX, 0.8,
label_text_color, 1)
# Measure overall FPS
total_processing_time = time.time() - start_processing_time
if total_processing_time == 0:
total_processing_time = 0.001 # handle zero division
total_fps = 1 / (total_processing_time)
# if FPS marking run time switch is turned on print some details on the image
if fps_marking:
label_text = f"FPS: {total_fps:.3}"
cv2.putText(batch, label_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX,
0.5, (0, 0, 255), 1)
label_text = f"Object detection inference time: {total_object_infer_time*1000:.4}ms"
cv2.putText(batch, label_text, (10, 50), cv2.FONT_HERSHEY_SIMPLEX,
0.5, (0, 0, 255), 1)
# Show the output image and save the output video
cv2.imshow('Frame', batch)
if args.save_output:
out.write(batch)
# Press q on keyboard to exit
# Press r on keyboard to toggle roll compensation
# Press u on keyboard to toggle ui drawings
# Press f on keyboard to fps drawings
ret = cv2.waitKey(20)
if ret & 0xFF == ord('q'):
break
elif ret & 0xFF == ord('u'):
ui_marking = not ui_marking
elif ret & 0xFF == ord('f'):
fps_marking = not fps_marking
# close the feed when stopping and finish the video saving
#feed.close()
if args.save_output:
out.release()
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 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()
