def main():
# model files check and download
check_and_download_models(WEIGHT_PATH, MODEL_PATH, REMOTE_PATH)
check_file_existance(FILE_PATH)
# prepare input data
canvas_3d = np.zeros((720, 1280, 3), dtype=np.uint8)
plotter = Plotter3d(canvas_3d.shape[:2])
canvas_3d_window_name = 'Canvas3D'
cv2.namedWindow(canvas_3d_window_name)
cv2.setMouseCallback(canvas_3d_window_name, Plotter3d.mouse_callback)
with open(FILE_PATH, 'r') as f:
extrinsics = json.load(f)
R = np.array(extrinsics['R'], dtype=np.float32)
t = np.array(extrinsics['t'], dtype=np.float32)
if args.video is None:
frame_provider = ImageReader([args.input])
is_video = False
else:
frame_provider = VideoReader(args.video)
is_video = True
fx = -1
delay = 1
esc_code = 27
p_code = 112
space_code = 32
mean_time = 0
img_mean = np.array([128, 128, 128], dtype=np.float32)
base_width_calculated = False
# net initialize
env_id = ailia.get_gpu_environment_id()
print(f'env_id: {env_id}')
net = ailia.Net(MODEL_PATH, WEIGHT_PATH, env_id=env_id)
# inference
for frame_id, frame in enumerate(frame_provider):
current_time = cv2.getTickCount()
if frame is None:
break
if not base_width_calculated:
IMAGE_WIDTH = frame.shape[1] * (IMAGE_HEIGHT / frame.shape[0])
IMAGE_WIDTH = int(IMAGE_WIDTH / STRIDE) * STRIDE
net.set_input_shape((1, 3, IMAGE_HEIGHT, IMAGE_WIDTH))
base_width_calculated = True
input_scale = IMAGE_HEIGHT / frame.shape[0]
scaled_img = cv2.resize(frame,
dsize=None,
fx=input_scale,
fy=input_scale)
# better to pad, but cut out for demo
scaled_img = scaled_img[:, 0:scaled_img.shape[1] -
(scaled_img.shape[1] % STRIDE)]
if fx < 0: # Focal length is unknown
fx = np.float32(0.8 * frame.shape[1])
normalized_img = (scaled_img.astype(np.float32) - img_mean) / 255.0
normalized_img = np.expand_dims(normalized_img.transpose(2, 0, 1),
axis=0)
# exectution
if is_video:
input_blobs = net.get_input_blob_list()
net.set_input_blob_data(normalized_img, input_blobs[0])
net.update()
features, heatmaps, pafs = net.get_results()
else:
print('Start inference...')
if args.benchmark:
print('BENCHMARK mode')
for i in range(5):
start = int(round(time.time() * 1000))
features, heatmaps, pafs = net.predict([normalized_img])
end = int(round(time.time() * 1000))
print(f'\tailia processing time {end - start} ms')
else:
features, heatmaps, pafs = net.predict([normalized_img])
inference_result = (features[-1].squeeze(), heatmaps[-1].squeeze(),
pafs[-1].squeeze())
poses_3d, poses_2d = parse_poses(inference_result, input_scale, STRIDE,
fx, is_video)
edges = []
if len(poses_3d):
poses_3d = rotate_poses(poses_3d, R, t)
poses_3d_copy = poses_3d.copy()
x = poses_3d_copy[:, 0::4]
y = poses_3d_copy[:, 1::4]
z = poses_3d_copy[:, 2::4]
poses_3d[:, 0::4], poses_3d[:, 1::4], poses_3d[:, 2::4] = -z, x, -y
poses_3d = poses_3d.reshape(poses_3d.shape[0], 19, -1)[:, :, 0:3]
edges = (Plotter3d.SKELETON_EDGES +
19 * np.arange(poses_3d.shape[0]).reshape(
(-1, 1, 1))).reshape((-1, 2))
plotter.plot(canvas_3d, poses_3d, edges)
if is_video:
cv2.imshow(canvas_3d_window_name, canvas_3d)
else:
cv2.imwrite(f'Canvas3D_{frame_id}.png', canvas_3d)
draw_poses(frame, poses_2d)
current_time = (cv2.getTickCount() -
current_time) / cv2.getTickFrequency()
if mean_time == 0:
mean_time = current_time
else:
mean_time = mean_time * 0.95 + current_time * 0.05
cv2.putText(frame, 'FPS: {}'.format(int(1 / mean_time * 10) / 10),
(40, 80), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255))
if is_video:
cv2.imshow('ICV 3D Human Pose Estimation', frame)
else:
cv2.imwrite(args.savepath, frame)
key = cv2.waitKey(delay)
if key == esc_code:
break
if key == p_code:
if delay == 1:
delay = 0
else:
delay = 1
if delay == 0 and args.rotate3d:
key = 0
while (key != p_code and key != esc_code and key != space_code):
plotter.plot(canvas_3d, poses_3d, edges)
cv2.imshow(canvas_3d_window_name, canvas_3d)
key = cv2.waitKey(33)
if key == esc_code:
break
else:
delay = 1
print('Script finished successfully.')
canvas_3d = np.zeros((720, 1280, 3), dtype=np.uint8)
plotter = Plotter3d(canvas_3d.shape[:2])
canvas_3d_window_name = 'Canvas 3D'
cv2.namedWindow(canvas_3d_window_name)
cv2.setMouseCallback(canvas_3d_window_name, Plotter3d.mouse_callback)
file_path = args.extrinsics_path
if file_path is None:
file_path = os.path.join('data', 'extrinsics.json')
with open(file_path, 'r') as f:
extrinsics = json.load(f)
R = np.array(extrinsics['R'], dtype=np.float32)
t = np.array(extrinsics['t'], dtype=np.float32)
frame_provider = ImageReader(args.images)
is_video = False
if args.video != '':
frame_provider = VideoReader(args.video)
is_video = True
base_height = args.height_size
fx = args.fx
delay = 1
esc_code = 27
p_code = 112
space_code = 32
mean_time = 0
for frame in frame_provider:
current_time = cv2.getTickCount()
if frame is None:
def run_inference(args):
from modules.inference_engine_pytorch import InferenceEnginePyTorch
socket_server = SocketServer(args.port)
joint_angle_calculator = JointAngleCalculator()
stride = 8
model_path = os.path.join('models', 'human-pose-estimation-3d.pth')
net = InferenceEnginePyTorch(model_path, "GPU")
canvas_3d = np.zeros((720, 1280, 3), dtype=np.uint8)
plotter = Plotter3d(canvas_3d.shape[:2])
canvas_3d_window_name = 'Canvas 3D'
cv2.namedWindow(canvas_3d_window_name)
cv2.setMouseCallback(canvas_3d_window_name, Plotter3d.mouse_callback)
file_path = None
if file_path is None:
file_path = os.path.join('data', 'extrinsics.json')
with open(file_path, 'r') as f:
extrinsics = json.load(f)
R = np.array(extrinsics['R'], dtype=np.float32)
t = np.array(extrinsics['t'], dtype=np.float32)
frame_provider = ImageReader(args.images)
is_video = False
if args.video != '':
frame_provider = VideoReader(args.video)
is_video = True
base_height = args.height_size
fx = 1 # focal length
delay = 1
esc_code = 27
p_code = 112
space_code = 32
mean_time = 0
for frame in frame_provider:
current_time = cv2.getTickCount()
if frame is None:
break
input_scale = base_height / frame.shape[0]
scaled_img = cv2.resize(frame, dsize=None, fx=input_scale, fy=input_scale)
scaled_img = scaled_img[:, 0:scaled_img.shape[1] - (scaled_img.shape[1] % stride)] # better to pad, but cut out for demo
if fx < 0: # Focal length is unknown
fx = np.float32(0.8 * frame.shape[1])
inference_result = net.infer(scaled_img)
poses_3d, poses_2d = parse_poses(inference_result, input_scale, stride, fx, is_video)
edges = []
if len(poses_3d):
poses_3d = rotate_poses(poses_3d, R, t)
poses_3d_copy = poses_3d.copy()
x = poses_3d_copy[:, 0::4]
y = poses_3d_copy[:, 1::4]
z = poses_3d_copy[:, 2::4]
poses_3d[:, 0::4], poses_3d[:, 1::4], poses_3d[:, 2::4] = -z, x, -y
poses_3d = poses_3d.reshape(poses_3d.shape[0], 19, -1)[:, :, 0:3]
edges = (Plotter3d.SKELETON_EDGES + 19 * np.arange(poses_3d.shape[0]).reshape((-1, 1, 1))).reshape((-1, 2))
plotter.plot(canvas_3d, poses_3d, edges)
cv2.imshow(canvas_3d_window_name, canvas_3d)
draw_poses(frame, poses_2d)
current_time = (cv2.getTickCount() - current_time) / cv2.getTickFrequency()
if mean_time == 0:
mean_time = current_time
else:
mean_time = mean_time * 0.95 + current_time * 0.05
cv2.putText(frame, 'FPS: {}'.format(int(1 / mean_time * 10) / 10),
(40, 80), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255))
cv2.imshow('ICV 3D Human Pose Estimation', frame)
key = cv2.waitKey(delay)
if key == esc_code:
break
if key == p_code:
if delay == 1:
delay = 0
else:
delay = 1
if delay == 0 or not is_video: # allow to rotate 3D canvas while on pause
key = 0
while (key != p_code
and key != esc_code
and key != space_code):
plotter.plot(canvas_3d, poses_3d, edges)
cv2.imshow(canvas_3d_window_name, canvas_3d)
key = cv2.waitKey(33)
if key == esc_code:
break
else:
delay = 1
joint_angles = joint_angle_calculator.calculate_angles(poses_3d)
if joint_angles:
socket_server.send_data(joint_angles)
canvas_3d = np.zeros((720, 1280, 3), dtype=np.uint8)
plotter = Plotter3d(canvas_3d.shape[:2])
canvas_3d_window_name = 'Canvas 3D'
cv2.namedWindow(canvas_3d_window_name)
cv2.setMouseCallback(canvas_3d_window_name, Plotter3d.mouse_callback)
file_path = args.extrinsics_path
if file_path is None:
file_path = os.path.join('data', 'extrinsics.json')
with open(file_path, 'r') as f:
extrinsics = json.load(f)
R = np.array(extrinsics['R'], dtype=np.float32)
t = np.array(extrinsics['t'], dtype=np.float32)
frame_provider = ImageReader(args.images)
is_video = False
if args.video != '':
frame_provider = VideoReader(args.video)
is_video = True
base_height = args.height_size
fx = args.fx
delay = 1
esc_code = 27
p_code = 112
space_code = 32
mean_time = 0
for frame in frame_provider:
current_time = cv2.getTickCount()
if frame is None:
def main():
parser = ArgumentParser(
description='Lightweight 3D human pose estimation demo. '
'Press esc to exit, "p" to (un)pause video or process next image.')
parser.add_argument(
'-m',
'--model',
help='Required. Path to checkpoint with a trained model '
'(or an .xml file in case of OpenVINO inference).',
type=str,
required=True)
parser.add_argument('--video',
help='Optional. Path to video file or camera id.',
type=str,
default='')
parser.add_argument('-o',
'--output',
help='output directory for estimated results',
default='./output')
parser.add_argument(
'-d',
'--device',
help='Optional. Specify the target device to infer on: CPU or GPU. '
'The demo will look for a suitable plugin for device specified '
'(by default, it is GPU).',
type=str,
default='GPU')
parser.add_argument(
'--use-openvino',
help='Optional. Run network with OpenVINO as inference engine. '
'CPU, GPU, FPGA, HDDL or MYRIAD devices are supported.',
action='store_true')
parser.add_argument(
'--use-tensorrt',
help='Optional. Run network with TensorRT as inference engine.',
action='store_true')
parser.add_argument('--images',
help='Optional. Path to input image(s).',
nargs='+',
default='')
parser.add_argument('--height-size',
help='Optional. Network input layer height size.',
type=int,
default=256)
parser.add_argument('--extrinsics-path',
help='Optional. Path to file with camera extrinsics.',
type=str,
default=None)
parser.add_argument('--fx',
type=np.float32,
default=-1,
help='Optional. Camera focal length.')
args = parser.parse_args()
if args.video == '' and args.images == '':
raise ValueError('Either --video or --image has to be provided')
infer_ctrl = InferCtrl(args.model,
args.height_size,
device=args.device,
openvino=args.use_openvino,
tensorrt=args.use_tensorrt,
extrinsics_path=args.extrinsics_path,
fx=args.fx)
frame_provider = ImageReader(args.images)
is_video = False
outname = args.images
if args.video != '':
frame_provider = VideoReader(args.video)
is_video = True
try:
cam_index = int(args.video)
outname = "output-cam-{}".format(cam_index)
except:
outname = ".".join(args.video.split(os.sep)[-1].split(".")[:-1])
fx = args.fx
mean_time = 0
i = 0
frame_size = None, None
dir_name = os.path.join(args.output, outname)
os.makedirs(dir_name, exist_ok=True)
outname_3d = os.path.join(dir_name, "{}-temp-3D.mp4".format(outname))
outname_frames = os.path.join(dir_name, "{}-temp.mp4".format(outname))
outname_combined_frames = os.path.join(
dir_name, "{}-combined-temp.mp4".format(outname))
outname_3d_compressed = os.path.join(dir_name, "{}-3D.mp4".format(outname))
outname_frames_compressed = os.path.join(dir_name,
"{}.mp4".format(outname))
outname_frames_combined_compressed = os.path.join(
dir_name, "{}-combined.mp4".format(outname))
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = None
out_3d = None
out_combined = None
fps = None
fps_out = 15
try:
results_json = []
for frame in frame_provider:
current_time = cv2.getTickCount()
if frame is None:
break
inference_result = infer_ctrl.infer(frame, is_video=False, fx=fx)
# print(inference_result)
# continue
# intepreting results
frame, canvas_3d = infer_ctrl.process_frame(
frame, inference_result)
combined_frame = infer_ctrl.process_frame(frame,
inference_result,
merged=True)
for key in inference_result.keys():
value = inference_result[key].get("value", [])
if "numpy" in str(type(value)):
inference_result[key]["value"] = getattr(
value, "tolist", lambda: value)()
results_json.append(inference_result)
current_time = (cv2.getTickCount() -
current_time) / cv2.getTickFrequency()
if mean_time == 0:
mean_time = current_time
else:
mean_time = mean_time * 0.95 + current_time * 0.05
fps = int(1 / mean_time * 10) / 10
cv2.putText(frame, 'processing FPS: {}'.format(fps), (40, 80),
cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255))
i += 1
if out is None or out_3d is None:
frame_size = tuple(int(i) for i in frame.shape[:2])
frame_size = frame_size[::-1]
frame_size_3d = tuple(int(i) for i in canvas_3d.shape)
frame_size_3d = canvas_3d.shape[1::-1]
frame_size_combined = tuple(
int(i) for i in combined_frame.shape)
frame_size_combined = combined_frame.shape[1::-1]
print(frame_size_3d, frame_size, frame_size_combined)
out_3d = cv2.VideoWriter(outname_3d, fourcc, fps_out,
frame_size_3d, True)
out = cv2.VideoWriter(outname_frames, fourcc, fps_out,
frame_size, True)
out_combined = cv2.VideoWriter(outname_combined_frames, fourcc,
fps_out, frame_size_combined,
True)
if out is not None:
out.write(frame)
if out_3d is not None:
out_3d.write(canvas_3d)
if out_combined is not None:
out_combined.write(combined_frame)
except KeyboardInterrupt:
print("[INFO] interrupted")
if out is not None:
out.release()
if out_3d is not None:
out_3d.release()
with open(os.path.join(dir_name, "results.json"), "w") as fp:
json.dump(results_json, fp)
fp.close()
try:
os.system(
f"ffmpeg -i {outname_frames} -loglevel error -vcodec libx264 {outname_frames_compressed}"
)
os.system(
f"ffmpeg -i {outname_3d} -loglevel error -vcodec libx264 {outname_3d_compressed}"
)
os.system(
f"ffmpeg -i {outname_combined_frames} -loglevel error -vcodec libx264 {outname_frames_combined_compressed}"
)
# os.system(f"rm -rf {outname_frames} {outname_3d}")
except:
traceback.print_exc()
print("[INFO] finished .... ")