Body v3 FAQ

This depends on the amount and type of sensors you want to use. You can find more info in the documentation included with the app. The easiest way to check whether your machine handles your sensor setup is to run the free trial.

Yes, definitely, the app was designed to handle data from different sensor types. Combining different sensor types can even be beneficial, since they all have different noise and accuracy characteristics in the pointcloud and skeleton data they deliver.

If you want to buy new, the Orbbec Femto Bolt is the practical answer: it uses the same depth technology as the Azure Kinect and is its closest current equivalent. Which of the others are still sold, and how they compare on resolution, field of view and range: depth sensor comparison, which also says what I would pick per job.

On quality, “best” is a subjective term and it depends. Azure Kinect and Kinect v2 produce the best tracking quality, but Azure Kinect generally needs a faster CPU/GPU and Kinect v2 is restricted to one sensor per machine (due to drivers/SDK). Kinect v1 has decent tracking and low GPU/USB requirements. Orbbec has ok tracking with low requirements and doesn’t need a power supply. RealSense needs no power supply but is generally very noisy.

The short answer is no. A depth sensor provides depth readings which the skeleton tracker can directly use for 3D joint placement. A webcam only delivers 2D data, and depth is estimated by the deep-learning based tracker.

Webcams are great extensions to a setup with one or more depth sensors, providing additional viewpoints and improving the quality of the solve. They are generally also less expensive, use less USB bandwidth, and some can provide 60 fps.

All webcams should work as long as they provide decent image quality. The tracker needs to know the field of view of the lens (in degrees) to estimate depth. Logitech webcams are automatically recognized and their field of view is known; for other brands you enter this value manually.

The software works with one or more sensors, as many as your hardware can handle. Additional sensors with different viewpoints can increase quality, because they see body parts that are occluded for another sensor. So basically the more the merrier.

This depends on the USB bandwidth of your machine, the type of sensors, and your CPU/GPU. Kinect v2 (Xbox One) is limited to a single sensor per machine by its driver/SDK, and Orbbec Astra to a single sensor when body tracking is enabled. Azure Kinect and Orbbec Femto sensors have no such limit and support hardware sync between them. For desktop machines you can add PCI-Express cards to expand USB bandwidth; a depth sensor generally needs its own USB host controller rather than just its own port, and most expansion cards share one controller across all their ports. You can also use additional machines with a network connection and use the sensors connected to them. See the documentation for more information.

Kinect v2 (Xbox One) has a driver/SDK limitation that prevents access to more than a single sensor per machine; there is no way around this. The reverse-engineered open source LibFreenect2 SDK could potentially access color/pointcloud data on some hardware setups, but that excludes body tracking.

Orbbec Astra can access multiple sensors for color/pointcloud data, but their body tracker only works on the first sensor at the moment.

This does not apply to the Orbbec Femto Bolt and Femto Mega, which run through the same code path as the Azure Kinect here: multiple sensors work, with wired hardware sync between them.

Generally speaking multiple sensors are always preferable, since they see more angles of the subject and have fewer occlusions. And newer sensor types almost always provide better, cleaner data than older ones, especially in the Kinect sensor range. Your particular setup I have most likely not tested, so if you need a more specific answer, test it with the trial and/or evaluation license.

No. One Multi Sensor license lets you connect as many sensors and machines as you want. When using multiple machines, the idea is to run the GUI on one machine and the headless/console version of the same app on your other machines, then use the Network Sensor option to receive data from the sensors on those machines.

Yes and no. Structured light sensors (like Kinect v1 and Orbbec Astra) can produce a bit more noise in overlapping areas. Kinect v2 (time of flight) can occasionally show some Z-wobble since those sensors cannot be synchronized. Azure Kinect can be synchronized fully, reducing any interference. In general, interference does not pose much of an issue for the solver.

Azure Kinect sensors have sync in/out ports on their backs (remove the cover) and can be daisy-chained using a simple 3.5mm audio jack cable. The app automatically detects this and sets things up accordingly. Internally the software synchronizes all incoming data using timestamps.

v3 supports aligning and fusing data from multiple sensors, which can improve quality with occlusions and/or increase the capture volume, depending on your setup. It has a new set of skeleton solvers compared to v1/v2, plus an optional deep-learning based tracker for improved quality and to help determine left/right/front/back of people. And v3 is in active development. See the Body family overview for a comparison.

The Deep-Tracking feature uses a deep-learning based tracker to improve quality by providing additional joint estimation in 2D and 3D. It is generally better at identifying left/right/front/back of people than some (older) sensors. It can run in GPU mode on an NVIDIA RTX 2000-series or newer GPU, or in CPU mode as long as your CPU supports AVX2 instructions (generally all CPUs from 2013 and later).

Yes, of course. You can find more info about updates and upgrades in the shop.

Data quality depends on how many sensors you use, their brand/type, how you set them up regarding angle and distance to the subject, and whether you use the deep tracker, for example. Due to these variances it’s best to try things out yourself for your particular setup, using the trial.

By default Body v3 is tuned for best body tracking performance and uses infrared streams where possible, as decoding those is generally faster than color streams and offers more stable lighting. You can switch to color using the top menu: Settings > Force using Color for video.

Many of the supported sensors don’t come with macOS drivers. On top of that, Apple dropped OpenGL support in favor of a proprietary substitute, doesn’t offer cross-compilation or NVIDIA CUDA support, requires a paid development subscription, and moved away from x86/x64 CPUs. Porting would mean rewriting major portions of the app for a very small user base, so there are currently no plans for a macOS release.

Some sensors offer Linux drivers and some portions of the app could potentially be ported to Linux. At the moment there are no plans, but that may change in the future if there is enough demand.