PrimeSense

PrimeSense
Information
Industry	3D sensing
Founded	2005
Headquarters	Tel Aviv, Israel (before acquisition)
Products	Depth-sensing chips and cameras; NiTE middleware
Parent	Apple (acquired in 2013)

PrimeSense was an Israeli developer of chips, camera reference designs and software for depth sensing. It supplied the 3D sensing technology and reference design used in the first generation of Microsoft's Kinect sensor for Xbox 360. The company was founded in late 2005 and had its headquarters in Tel Aviv before Apple acquired it in November 2013.

PrimeSense hardware also provided depth images for research in augmented reality and virtual reality. Examples include KinectFusion's reconstruction of physical scenes for AR and Kinect-based body tracking in a virtual rehabilitation experiment. These systems combined depth measurements with software developed by researchers or Microsoft.

Reviewed 12 October 2026. Historical company and acquisition; documented depth-sensing products, OpenNI and NiTE; cited XR experiments and software preservation resources. About review dates.

Company history

The Weizmann Institute's 2014 profile of founder Aviad Maizels dates PrimeSense's establishment to late 2005. The profile describes the company's initial interest in making games respond to people and their surroundings through visual input. In the joint Microsoft and PrimeSense announcement of March 2010, Maizels was president and founder, and Inon Beracha was chief executive officer.

Date	Development
Late 2005	PrimeSense was founded.
31 March 2010	Microsoft and PrimeSense announced that PrimeSense would supply 3D sensing technology for Project Natal, the Xbox 360 sensor then in development. Microsoft's hardware engineering teams adapted the PrimeSensor reference design for the product.
December 2010	PrimeSense released sensor drivers and an OpenNI API as open source, alongside Windows and Linux binaries for its NITE tracking library.
11 December 2012	PrimeSense announced the Capri embedded depth-sensing reference design and system on chip, with demonstrations planned for CES 2013.
November 2013	Apple and PrimeSense confirmed the acquisition. Apple did not disclose the price or its intended use of the technology to Reuters.
April 2014	After the original OpenNI website closed, Occipital provided a separate resource for OpenNI 2 source code, binaries and documentation.

Depth-sensing technology

PrimeSense called its approach Light Coding. Its product brochure describes projecting near-infrared light into a scene, observing it with a CMOS image sensor, and processing the returned pattern on a dedicated system on chip to produce a depth image. The depth calculation took place on the chip before transmission to the host computer over USB.

The first Kinect used structured light and triangulation. An infrared camera observed a projected pattern and compared it with a reference pattern captured at a known distance. The displacement of the pattern supplied disparity measurements from which the system calculated depth. A separate RGB camera captured color images; calibration of the two cameras allowed color to be associated with the measured 3D points.

A PrimeSense patent application published in May 2010, Depth mapping using projected patterns, describes reconstructing a 3D surface map from offsets between a projected spot pattern and a reference image. It also describes using the resulting measurements of body movement for gesture input. The named inventors are Barak Freedman, Alexander Shpunt, Meir Machline and Yoel Arieli.

Kinect generations

The first Kinect's PrimeSense-based structured-light measurement differs from the later Kinect for Xbox One/Kinect v2, which uses time of flight. Sarbolandi, Lefloch and Kolb's 2015 comparison evaluates the two generations separately because their measurement principles produce different error characteristics.

PrimeSense supplied a sensing component and reference design for the original product. Microsoft's March 2010 announcement identifies Microsoft's Xbox hardware team as the developer of the adapted sensor. Body-pose estimation is a further software operation: the 2015 virtual rehabilitation study, for example, used Kinect for Windows SDK 1.7 for skeleton tracking.

Products and implementations

PrimeSense supplied both components for device makers and complete reference or evaluation sensors. The following names describe its historical product line; the specifications are those of the cited contemporary documents.

Product or component	Documented scope
PS1080 system on chip and PrimeSensor reference design	Listed in the March 2010 Microsoft/PrimeSense announcement. The reference design formed the basis of Microsoft's Project Natal sensor.
PrimeSensor Development Kit 5.0 (PSDK 5.0)	A USB-powered development sensor offered to OpenNI developers in December 2010. Willow Garage described it as smaller and lighter than Kinect.
Carmine 1.08	Depth-sensing reference design with a specified operating range of 0.8-3.5 m and a 640 x 480 depth image in the company brochure.
Carmine 1.09	Evaluation sensor for shorter distances, with a specified operating range of 0.35-1.4 m and a 640 x 480 depth image in the company brochure.
Capri / Capri 1.25	A smaller reference design intended for integration into consumer devices. The December 2012 announcement scheduled development samples for leading manufacturers for mid-2013; the brochure names the Capri PS1200 system on chip.
NITE / NiTE	Middleware for interpreting depth data, including user and hand tracking. PrimeSense's April 2013 migration guide describes NiTE as proprietary software with its own license.

An independent 2016 study tested the original Kinect, ASUS Xtion and Occipital Structure Sensor as three implementations of PrimeSense sensing technology. It analyzed their depth errors rather than treating shared technology as a guarantee of identical performance.

OpenNI and NiTE

Sensor access and tracking algorithms

In December 2010, PrimeSense released open-source sensor drivers and an OpenNI API that gave applications a common interface to RGB-D sensors. It separately distributed binary versions of NITE for skeleton and handpoint tracking. Willow Garage worked with PrimeSense on the drivers and their integration with ROS.

The surviving PrimeSense Sensor repository documents version 5.1.2.1, dated 15 May 2012, and build instructions that require OpenNI 1.5.x.x. Its source is distributed under the Apache 2.0 license. These are requirements for that historical driver version.

PrimeSense changed the architecture in OpenNI 2. Its April 2013 migration guide describes OpenNI 2 as an interface for depth and image data through device drivers, with NiTE in a separate middleware package and API. The guide distinguishes open-source, Apache-licensed OpenNI from proprietary NiTE. Access to an OpenNI driver therefore did not by itself provide the NiTE tracking algorithms.

The OpenNI 2 programmer's guide describes opening a physical device or a recorded ONI file, creating video streams, and obtaining frames. It also provides recording and playback interfaces. This lets an application handle recorded sensor input through the same device and stream abstractions used for hardware.

NiTE 2 tracking output

NiTE 2.0's HandTracker API returns positions of hands and detected gestures for individual depth frames. Its documentation describes tracking hand points for gesture-based interfaces, including seated users whose whole body is not visible. This API depends on an initialized OpenNI 2.0 device supplying depth data.

The UserTracker API provides scene segmentation, skeleton tracking, floor-plane detection and pose detection. It assigns tracked users identifiers, but the documentation warns that an identifier can change when tracking is lost or a user leaves the field of view. Persistent recognition of a person after tracking loss requires application logic.

Preservation after the acquisition

In April 2014, Occipital stated that it had preserved OpenNI 2 source code, binaries and documentation after openni.org closed, and intended to support other PrimeSense-based sensors, including Carmine and ASUS Xtion. The original OpenNI/OpenNI2 repository now explicitly states that it is no longer maintained and directs users to volunteer-maintained alternatives.

Structure's resource page retains OpenNI 2.2.0.33 Beta binaries and documentation for the original Structure Sensor. It also specifies limitations, including Intel-only macOS binaries, no iOS support in OpenNI 2, and no OpenNI support for Structure Core. These resources describe particular sensors and software versions.

Uses in VR, AR and 3D capture

Scene reconstruction and AR interaction

KinectFusion, presented at UIST 2011, used depth data from a moving Kinect to track the camera and combine successive views into a surface model. The researchers demonstrated virtual objects aligned with the captured scene, occlusion of virtual graphics by real geometry, and simulated particles interacting with reconstructed surfaces. The paper also describes extending the reconstruction process to separate foreground interaction from the static background.

This work connected depth acquisition with spatial registration and interaction. Its reconstructed models could be imported into CAD or other 3D modeling software, while its AR demonstrations used the reconstructed geometry to combine virtual and physical content.

Body tracking in a virtual environment

A 2015 study compared optical, electromagnetic and Kinect skeleton tracking in a virtual rehabilitation setup. A television displayed a virtual stepping task, and the system represented the participants' tracked feet in that environment. The Kinect implementation used Microsoft's SDK 1.7. The study measured tracking accuracy and jitter and collected feedback from healthy participants, stroke survivors and therapists; it evaluated the three tracking configurations in that particular task.

Calibration and object reconstruction

Carmine 1.09 also appears in a 2017 experiment comparing calibration methods for Kinect v1, Carmine 1.09 and Kinect v2. The researchers applied calibrated measurements to reconstruction of physical objects from multiple views. Calibration changed the accuracy of the resulting reconstructions, with the methods performing differently across the sensors and tests.

Measurement limits

The depth image and the color image require geometric alignment. Khoshelham and Oude Elberink's 2012 Kinect experiment found that stereo calibration reduced misalignment and systematic errors. Their plane measurements also showed increasing depth noise and coarser depth resolution as distance increased. Strong illumination reduced pattern contrast, while shadows, occlusion and shiny surfaces could leave gaps in the point cloud. These findings concern the first-generation Kinect and the study's measurement conditions.

Sarbolandi and colleagues' comparison examined further effects, including interference between overlapping active cameras and changes during device warm-up. Their results depended on sensor generation and experimental arrangement. Using several cameras to capture a person or a room therefore requires accounting for their illumination and measurement behavior.

The 2016 study of low-cost range sensors similarly distinguished systematic measurement error from random noise. It found usable applications in shape acquisition while identifying limits for precision inspection, including thermal stability and measurement uncertainty.