Live 3D Human Pose Estimation with OpenVINO

This tutorial is also available as a Jupyter notebook that can be cloned directly from GitHub. See the installation guide for instructions to run this tutorial locally on Windows, Linux or macOS. To run without installing anything, click the “launch binder” button.

Binder Github

This notebook demonstrates live 3D Human Pose Estimation with OpenVINO via a webcam. We utilize the model human-pose-estimation-3d-0001 from Open Model Zoo. At the end of this notebook, you will see live inference results from your webcam (if available). Alternatively, you can also upload a video file to test out the algorithms. Make sure you have properly installed theJupyter extensionand been using JupyterLab to run the demo as suggested in the ````


To use a webcam, you must run this Jupyter notebook on a computer with a webcam. If you run on a remote server, the webcam will not work. However, you can still do inference on a video file in the final step. This demo utilizes the Python interface in Three.js integrated with WebGL to process data from the model inference. These results are processed and displayed in the notebook.

To ensure that the results are displayed correctly, run the code in a recommended browser on one of the following operating systems: Ubuntu, Windows: Chrome, macOS: Safari.

Table of contents:



The pythreejs extension may not display properly when using the latest Jupyter Notebook release (2.4.1). Therefore, it is recommended to use Jupyter Lab instead.

!pip install pythreejs
Collecting pythreejs
  Using cached pythreejs-2.4.2-py3-none-any.whl (3.4 MB)
Requirement already satisfied: ipywidgets>=7.2.1 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from pythreejs) (8.1.0)
Collecting ipydatawidgets>=1.1.1 (from pythreejs)
  Obtaining dependency information for ipydatawidgets>=1.1.1 from
  Using cached ipydatawidgets-4.3.5-py2.py3-none-any.whl.metadata (1.4 kB)
Requirement already satisfied: numpy in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from pythreejs) (1.23.5)
Requirement already satisfied: traitlets in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from pythreejs) (5.9.0)
Collecting traittypes>=0.2.0 (from ipydatawidgets>=1.1.1->pythreejs)
  Using cached traittypes-0.2.1-py2.py3-none-any.whl (8.6 kB)
Requirement already satisfied: comm>=0.1.3 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipywidgets>=7.2.1->pythreejs) (0.1.4)
Requirement already satisfied: ipython>=6.1.0 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipywidgets>=7.2.1->pythreejs) (8.12.2)
Requirement already satisfied: widgetsnbextension~=4.0.7 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipywidgets>=7.2.1->pythreejs) (4.0.8)
Requirement already satisfied: jupyterlab-widgets~=3.0.7 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipywidgets>=7.2.1->pythreejs) (3.0.8)
Requirement already satisfied: backcall in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (0.2.0)
Requirement already satisfied: decorator in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (4.4.2)
Requirement already satisfied: jedi>=0.16 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (0.19.0)
Requirement already satisfied: matplotlib-inline in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (0.1.6)
Requirement already satisfied: pickleshare in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (0.7.5)
Requirement already satisfied: prompt-toolkit!=3.0.37,<3.1.0,>=3.0.30 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (3.0.39)
Requirement already satisfied: pygments>=2.4.0 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (2.16.1)
Requirement already satisfied: stack-data in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (0.6.2)
Requirement already satisfied: typing-extensions in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (4.7.1)
Requirement already satisfied: pexpect>4.3 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (4.8.0)
Requirement already satisfied: parso<0.9.0,>=0.8.3 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from jedi>=0.16->ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (0.8.3)
Requirement already satisfied: ptyprocess>=0.5 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from pexpect>4.3->ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (0.7.0)
Requirement already satisfied: wcwidth in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from prompt-toolkit!=3.0.37,<3.1.0,>=3.0.30->ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (0.2.6)
Requirement already satisfied: executing>=1.2.0 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from stack-data->ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (1.2.0)
Requirement already satisfied: asttokens>=2.1.0 in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from stack-data->ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (2.2.1)
Requirement already satisfied: pure-eval in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from stack-data->ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (0.2.2)
Requirement already satisfied: six in /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages (from asttokens>=2.1.0->stack-data->ipython>=6.1.0->ipywidgets>=7.2.1->pythreejs) (1.16.0)
Using cached ipydatawidgets-4.3.5-py2.py3-none-any.whl (271 kB)
DEPRECATION: pytorch-lightning 1.6.5 has a non-standard dependency specifier torch>=1.8.*. pip 23.3 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of pytorch-lightning or contact the author to suggest that they release a version with a conforming dependency specifiers. Discussion can be found at
Installing collected packages: traittypes, ipydatawidgets, pythreejs
Successfully installed ipydatawidgets-4.3.5 pythreejs-2.4.2 traittypes-0.2.1


import collections
import sys
import time
from pathlib import Path

import cv2
import ipywidgets as widgets
import numpy as np
from IPython.display import clear_output, display
from openvino.runtime import Core

import notebook_utils as utils

import engine.engine3js as engine
from engine.parse_poses import parse_poses

The model

Download the model

We use omz_downloader, which is a command line tool from the openvino-dev package. omz_downloader automatically creates a directory structure and downloads the selected model.

# directory where model will be downloaded
base_model_dir = "model"

# model name as named in Open Model Zoo
model_name = "human-pose-estimation-3d-0001"
# selected precision (FP32, FP16)
precision = "FP32"

BASE_MODEL_NAME = f"{base_model_dir}/public/{model_name}/{model_name}"
model_path = Path(BASE_MODEL_NAME).with_suffix(".pth")
onnx_path = Path(BASE_MODEL_NAME).with_suffix(".onnx")

ir_model_path = f"model/public/{model_name}/{precision}/{model_name}.xml"
model_weights_path = f"model/public/{model_name}/{precision}/{model_name}.bin"

if not model_path.exists():
    download_command = (
        f"omz_downloader " f"--name {model_name} " f"--output_dir {base_model_dir}"
    ! $download_command
################|| Downloading human-pose-estimation-3d-0001 ||################

========== Downloading model/public/human-pose-estimation-3d-0001/human-pose-estimation-3d-0001.tar.gz

========== Unpacking model/public/human-pose-estimation-3d-0001/human-pose-estimation-3d-0001.tar.gz

Convert Model to OpenVINO IR format

The selected model comes from the public directory, which means it must be converted into OpenVINO Intermediate Representation (OpenVINO IR). We use omz_converter to convert the ONNX format model to the OpenVINO IR format.

if not onnx_path.exists():
    convert_command = (
        f"omz_converter "
        f"--name {model_name} "
        f"--precisions {precision} "
        f"--download_dir {base_model_dir} "
        f"--output_dir {base_model_dir}"
    ! $convert_command
========== Converting human-pose-estimation-3d-0001 to ONNX
Conversion to ONNX command: /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/bin/python -- /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/openvino/model_zoo/internal_scripts/ --model-path=model/public/human-pose-estimation-3d-0001 --model-name=PoseEstimationWithMobileNet --model-param=is_convertible_by_mo=True --import-module=model --weights=model/public/human-pose-estimation-3d-0001/human-pose-estimation-3d-0001.pth --input-shape=1,3,256,448 --input-names=data --output-names=features,heatmaps,pafs --output-file=model/public/human-pose-estimation-3d-0001/human-pose-estimation-3d-0001.onnx

ONNX check passed successfully.

========== Converting human-pose-estimation-3d-0001 to IR (FP32)
Conversion command: /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/bin/python -- /opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-475/.workspace/scm/ov-notebook/.venv/bin/mo --framework=onnx --output_dir=/tmp/tmpgwxi10io --model_name=human-pose-estimation-3d-0001 --input=data '--mean_values=data[128.0,128.0,128.0]' '--scale_values=data[255.0,255.0,255.0]' --output=features,heatmaps,pafs --input_model=model/public/human-pose-estimation-3d-0001/human-pose-estimation-3d-0001.onnx '--layout=data(NCHW)' '--input_shape=[1, 3, 256, 448]' --compress_to_fp16=False

[ INFO ] The model was converted to IR v11, the latest model format that corresponds to the source DL framework input/output format. While IR v11 is backwards compatible with OpenVINO Inference Engine API v1.0, please use API v2.0 (as of 2022.1) to take advantage of the latest improvements in IR v11.
Find more information about API v2.0 and IR v11 at
[ SUCCESS ] Generated IR version 11 model.
[ SUCCESS ] XML file: /tmp/tmpgwxi10io/human-pose-estimation-3d-0001.xml
[ SUCCESS ] BIN file: /tmp/tmpgwxi10io/human-pose-estimation-3d-0001.bin

Select inference device

Select device from dropdown list for running inference using OpenVINO:

core = Core()

device = widgets.Dropdown(
    options=core.available_devices + ["AUTO"],

Dropdown(description='Device:', index=1, options=('CPU', 'AUTO'), value='AUTO')

Load the model

Converted models are located in a fixed structure, which indicates vendor, model name and precision.

First, initialize the inference engine, OpenVINO Runtime. Then, read the network architecture and model weights from the .bin and .xml files to compile for the desired device. An inference request is then created to infer the compiled model.

# initialize inference engine
ie_core = Core()
# read the network and corresponding weights from file
model = ie_core.read_model(model=ir_model_path, weights=model_weights_path)
# load the model on the specified device
compiled_model = ie_core.compile_model(model=model, device_name=device.value)
infer_request = compiled_model.create_infer_request()
input_tensor_name = model.inputs[0].get_any_name()

# get input and output names of nodes
input_layer = compiled_model.input(0)
output_layers = list(compiled_model.outputs)

The input for the model is data from the input image and the outputs are heat maps, PAF (part affinity fields) and features.

input_layer.any_name, [o.any_name for o in output_layers]
('data', ['features', 'heatmaps', 'pafs'])


Model Inference

Frames captured from video files or the live webcam are used as the input for the 3D model. This is how you obtain the output heat maps, PAF (part affinity fields) and features.

def model_infer(scaled_img, stride):
    Run model inference on the input image

        scaled_img: resized image according to the input size of the model
        stride: int, the stride of the window

    # Remove excess space from the picture
    img = scaled_img[
        0 : scaled_img.shape[0] - (scaled_img.shape[0] % stride),
        0 : scaled_img.shape[1] - (scaled_img.shape[1] % stride),

    img = np.transpose(img, (2, 0, 1))[
    infer_request.infer({input_tensor_name: img})
    # A set of three inference results is obtained
    results = {
        name: infer_request.get_tensor(name).data[:]
        for name in {"features", "heatmaps", "pafs"}
    # Get the results
    results = (results["features"][0], results["heatmaps"][0], results["pafs"][0])

    return results

Draw 2D Pose Overlays

We need to define some connections between the joints in advance, so that we can draw the structure of the human body in the resulting image after obtaining the inference results. Joints are drawn as circles and limbs are drawn as lines. The code is based on the 3D Human Pose Estimation Demo from Open Model Zoo.

# 3D edge index array
body_edges = np.array(
        [0, 1],
        [0, 9], [9, 10], [10, 11],    # neck - r_shoulder - r_elbow - r_wrist
        [0, 3], [3, 4], [4, 5],       # neck - l_shoulder - l_elbow - l_wrist
        [1, 15], [15, 16],            # nose - l_eye - l_ear
        [1, 17], [17, 18],            # nose - r_eye - r_ear
        [0, 6], [6, 7], [7, 8],       # neck - l_hip - l_knee - l_ankle
        [0, 12], [12, 13], [13, 14],  # neck - r_hip - r_knee - r_ankle

body_edges_2d = np.array(
        [0, 1],                       # neck - nose
        [1, 16], [16, 18],            # nose - l_eye - l_ear
        [1, 15], [15, 17],            # nose - r_eye - r_ear
        [0, 3], [3, 4], [4, 5],       # neck - l_shoulder - l_elbow - l_wrist
        [0, 9], [9, 10], [10, 11],    # neck - r_shoulder - r_elbow - r_wrist
        [0, 6], [6, 7], [7, 8],       # neck - l_hip - l_knee - l_ankle
        [0, 12], [12, 13], [13, 14],  # neck - r_hip - r_knee - r_ankle

def draw_poses(frame, poses_2d, scaled_img, use_popup):
    Draw 2D pose overlays on the image to visualize estimated poses.
    Joints are drawn as circles and limbs are drawn as lines.

    :param frame: the input image
    :param poses_2d: array of human joint pairs
    for pose in poses_2d:
        pose = np.array(pose[0:-1]).reshape((-1, 3)).transpose()
        was_found = pose[2] > 0

        pose[0], pose[1] = (
            pose[0] * frame.shape[1] / scaled_img.shape[1],
            pose[1] * frame.shape[0] / scaled_img.shape[0],

        # Draw joints.
        for edge in body_edges_2d:
            if was_found[edge[0]] and was_found[edge[1]]:
                    tuple(pose[0:2, edge[0]].astype(np.int32)),
                    tuple(pose[0:2, edge[1]].astype(np.int32)),
                    (255, 255, 0),
        # Draw limbs.
        for kpt_id in range(pose.shape[1]):
            if pose[2, kpt_id] != -1:
                    tuple(pose[0:2, kpt_id].astype(np.int32)),
                    (0, 255, 255),

    return frame

Main Processing Function

Run 3D pose estimation on the specified source. It could be either a webcam feed or a video file.

def run_pose_estimation(source=0, flip=False, use_popup=False, skip_frames=0):
    2D image as input, using OpenVINO as inference backend,
    get joints 3D coordinates, and draw 3D human skeleton in the scene

    :param source:      The webcam number to feed the video stream with primary webcam set to "0", or the video path.
    :param flip:        To be used by VideoPlayer function for flipping capture image.
    :param use_popup:   False for showing encoded frames over this notebook, True for creating a popup window.
    :param skip_frames: Number of frames to skip at the beginning of the video.

    focal_length = -1  # default
    stride = 8
    player = None
    skeleton_set = None

        # create video player to play with target fps  video_path
        # get the frame from camera
        # You can skip first N frames to fast forward video. change 'skip_first_frames'
        player = utils.VideoPlayer(source, flip=flip, fps=30, skip_first_frames=skip_frames)
        # start capturing

        input_image =
        # set the window size
        resize_scale = 450 / input_image.shape[1]
        windows_width = int(input_image.shape[1] * resize_scale)
        windows_height = int(input_image.shape[0] * resize_scale)

        # use visualization library
        engine3D = engine.Engine3js(grid=True, axis=True, view_width=windows_width, view_height=windows_height)

        if use_popup:
            # display the 3D human pose in this notebook, and origin frame in popup window
            title = "Press ESC to Exit"
            cv2.namedWindow(title, cv2.WINDOW_KEEPRATIO | cv2.WINDOW_AUTOSIZE)
            # set the 2D image box, show both human pose and image in the notebook
            imgbox = widgets.Image(
                format="jpg", height=windows_height, width=windows_width
            display(widgets.HBox([engine3D.renderer, imgbox]))

        skeleton = engine.Skeleton(body_edges=body_edges)

        processing_times = collections.deque()

        while True:
            # grab the frame
            frame =
            if frame is None:
                print("Source ended")

            # resize image and change dims to fit neural network input
            # (see
            scaled_img = cv2.resize(frame, dsize=(model.inputs[0].shape[3], model.inputs[0].shape[2]))

            if focal_length < 0:  # Focal length is unknown
                focal_length = np.float32(0.8 * scaled_img.shape[1])

            # inference start
            start_time = time.time()
            # get results
            inference_result = model_infer(scaled_img, stride)

            # inference stop
            stop_time = time.time()
            processing_times.append(stop_time - start_time)
            # Process the point to point coordinates of the data
            poses_3d, poses_2d = parse_poses(inference_result, 1, stride, focal_length, True)

            # use processing times from last 200 frames
            if len(processing_times) > 200:

            processing_time = np.mean(processing_times) * 1000
            fps = 1000 / processing_time

            if len(poses_3d) > 0:
                # From here, you can rotate the 3D point positions using the function "draw_poses",
                # or you can directly make the correct mapping below to properly display the object image on the screen
                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 + np.ones(poses_3d[:, 2::4].shape) * 200,
                    -y + np.ones(poses_3d[:, 2::4].shape) * 100,

                poses_3d = poses_3d.reshape(poses_3d.shape[0], 19, -1)[:, :, 0:3]
                people = skeleton(poses_3d=poses_3d)

                except Exception:

                skeleton_set = people

                # draw 2D
                frame = draw_poses(frame, poses_2d, scaled_img, use_popup)

                    skeleton_set = None
                except Exception:

                f"Inference time: {processing_time:.1f}ms ({fps:.1f} FPS)",
                (10, 30),
                (0, 0, 255),

            if use_popup:
                cv2.imshow(title, frame)
                key = cv2.waitKey(1)
                # escape = 27, use ESC to exit
                if key == 27:
                # encode numpy array to jpg
                imgbox.value = cv2.imencode(
                    params=[cv2.IMWRITE_JPEG_QUALITY, 90],


    except KeyboardInterrupt:
    except RuntimeError as e:
        if player is not None:
            # stop capturing
        if use_popup:
        if skeleton_set:


Run Live Pose Estimation

Run, using a webcam as the video input. By default, the primary webcam is set with source=0. If you have multiple webcams, each one will be assigned a consecutive number starting at 0. Set flip=True when using a front-facing camera. Some web browsers, especially Mozilla Firefox, may cause flickering. If you experience flickering, set use_popup=True.


1. To use this notebook with a webcam, you need to run the notebook on a computer with a webcam. If you run the notebook on a server (e.g. Binder), the webcam will not work.

2. Popup mode may not work if you run this notebook on a remote computer (e.g. Binder).

Using the following method, you can click and move your mouse over the picture on the left to interact.

run_pose_estimation(source=0, flip=True, use_popup=False)

Run Pose Estimation on a Video File

If you do not have a webcam, you can still run this demo with a video file. Any format supported by OpenCV will work.

You can click and move your mouse over the picture on the left to interact.

# video url
video_path = ""
run_pose_estimation(source=video_path, flip=False, use_popup=False, skip_frames=10)