Set up Body Pose Estimation (BPE)

Body Pose Estimation allows an accurate live estimation of the pose and position of a human body directly from a video image, using an AI powered framework from NVIDIA. It provides a real-time set of 3D data points corresponding to a skeleton with bones and body parts.

With it you can:

• position the VS compositing plane, allowing the talent on the green screen set to freely move towards and away from the camera

• render shadows and reflections more accurately and even render footprints

• allow the talent to interact with the virtual environment using

  • physics

  • trigger boxes

  • hand actions

  • or a clicker control for pick up & place logic

Limitations for the current NVIDIA AR SDK

• Only a single person is allowed within the camera field of view (including humanoids like mannequins)

  • if a full person cannot be seen, it will try to infer and extrapolate bone positions

• The resolution of the fingers is poor or erratic, so hand poses cannot be determined

• Real-time zooming or focussing is not supported, unless you’re sharing Body Pose Estimation data from another machine.

• In complex scenes, the GPU processing overhead can be high and could affect rendering performance

  • See Share Body Pose Estimation with other engines below to overcome this

• It is likely that the skeleton position is always one frame behind the actual position

• Only NVIDIA GeForce RTX 20 and 30 Series or NVIDIA RTX professional GPUs are supported

  • Use NVIDIA GPUs with Tensor Cores, ideally Tensor 2 Cores

Install NVIDIA AR SDK

The latest Nvidia AR SDK 0.8.x is NOT compatible with Pixotope just yet. Please use Nvidia AR SDK 0.7.6 which you can download here:

• For NVIDIA GeForce RTX 30 Series or NVIDIA RTX professional GPUs
  https://international.download.nvidia.com/Windows/broadcast/sdk/0.7.6/nvidia_ar_sdk_installer_ampere.exe

• For NVIDIA GeForce RTX 20 Series or NVIDIA Quadro RTX
  https://international.download.nvidia.com/Windows/broadcast/sdk/0.7.6/nvidia_ar_sdk_installer_turing.exe

1. Install the AR SDK from
   https://www.nvidia.com/en-us/geforce/broadcasting/broadcast-sdk/resources/
   

   

2. Restart Pixotope

Only NVIDIA GeForce RTX 20 and 30 Series or NVIDIA RTX professional GPUs are supported

Enable Body Pose Estimation

In Director

1. Go to SETUP > Configure > Routing

2. Enable Body Pose Estimation for the video input which should be analyzed

Use Body Pose Estimation in level

In Editor

1. Go to Place Actors > Pixotope

2. Add the "BPE & Plane" actor to the level

   • Dropping this actor into the scene creates:

     • a "BPE Compositing Mesh"

     • and a "VS Internal Compositing Plane" which follows the "BPE Compositing Mesh" in position and scale

       • Rotate with Camera is enabled

3. Choose if the plane should auto-size with "Use Body Pose Estimation for scale"

   • TPose - always provide enough space for handling fast movement and hand action covering the maximum extent of the body

   • Tight - closely follow the detected body with a small margin around to ensure it doesn’t clip (for smaller greenscreens)

   • Off - no autosizing, manual sizing

Learn more about the BPE Compositing Mesh

The same setup as the BPE & Plane actor can be achieved by adding

• a "BPE Compositing Mesh"

• a "VS Internal Compositing Plane"

  • with "Use body pose estimation for position" enabled

• and linking "Position Scale" of both actors

Share Body Pose Estimation with other engines

It can be useful to share the Body Pose Estimation data with other engines if:

1. The rendering overhead is too high to generate Body Pose Estimation on the same machine rendering the scene

2. The main camera is moving in too tight on the presenter, and the Body Pose Estimation cannot resolve the skeleton

3. You are in a multi-cam environment where multiple cameras need to use the Body Pose Estimation data

4. There is a need to perform dynamic zoom and focus in camera

Scenario 1 - Single Camera, Second Engine

Use a separate Pixotope engine with a camera input from the main camera to generate the Body Pose Estimation data.

On the source engine - generates the Body Pose Estimation data

1. Open the "Utilities" panel and go to BPE > Export

2. Click "Start Export"

   • this stores the Body Pose Estimation data in the Store and makes it accessible for all connected engines

On the target engines - uses the generated data

1. Open the "Utilities" panel and go to BPE > Import

2. Click "Start Import"

To assure that the importing/target machine does not accidentally also generate BPE data (which would cause the doubling of data) we will disable BPE on the target machine.

The Body Pose Estimation data will be 1 frame delayed from real-time.

The target engines do not need to have the same high powered GPU to use the Body Pose Estimation data.

Scenario 2 - Witness Camera, Second Engine

Use a separate camera to provide the images for the Body Pose Estimation source engine.

1. Set up witness camera and engine

   • The “witness” camera should

     • be setup to view the talent on screen wherever they are on the green screen set

     • be free from occlusions from other camera’s motion or set elements

     • not capture any people to the edge or off-set

   • the engine creating the Body Pose Estimation requires a high performing GPU - the target systems do not

2. Set up Body Pose Estimation sharing as shown in Scenario 1

Advanced

Use Body Pose Estimation in blueprints

All Body Pose Estimation points can be directly accessed in blueprints via the BPE Manager.

In addition, the Procedural Composite Mesh has multiple “attachment” points that can be used.