POSIT

POSIT
Name	POSIT
Developer	David R. Martin
Released	1987
Programming language	C, C++
Operating system	Cross-platform
Genre	Computer vision, Pose estimation

POSIT

POSIT is an algorithm for estimating three-dimensional pose from two-dimensional images using correspondences between image points and model points. It combines geometric reasoning with iterative refinement to recover orientation and position relative to a calibrated camera, and it has been influential in computer vision, robotics, augmented reality and photogrammetry. The method is historically associated with early pose-estimation work and has been cited in literature alongside methods developed at institutions such as Massachusetts Institute of Technology, Carnegie Mellon University, Stanford University, California Institute of Technology, and Imperial College London.

Overview

POSIT was introduced to solve the pose problem given a set of known three-dimensional model points and their two-dimensional projections in an image. The algorithm bridges ideas from projective geometry used in research at University of Oxford, University of Cambridge, ETH Zurich, University of Toronto, and University of California, Berkeley. It exploits correspondences similar to approaches found in work by researchers at MIT Media Lab, Bell Labs, AT&T Labs, Siemens AG, and Nokia Research Center. POSIT’s development was contemporaneous with landmark projects from NASA, European Space Agency, Roscosmos, JAXA, and practical deployments in systems produced by Sony Corporation, Panasonic Corporation, Intel Corporation, and IBM research groups.

The method assumes a pinhole camera model used in studies at Tokyo Institute of Technology, Korea Advanced Institute of Science and Technology, Tsinghua University, Peking University, and National University of Singapore. It relates to calibration techniques promoted by Microsoft Research, Google Research, Facebook AI Research, and datasets popularized by Stanford Vision and Learning Lab and Oxford Visual Geometry Group.

Algorithm

POSIT computes pose by iteratively solving for scaling factors and rotation using a decomposition that resembles algorithmic ideas from work at Bell Labs Research, University of Illinois Urbana-Champaign, Princeton University, Yale University, and Cornell University. Starting with an initial orthographic approximation akin to methods in publications from University College London, Delft University of Technology, KTH Royal Institute of Technology, and Seoul National University, POSIT refines estimates using an Alternating Least Squares style update comparable to optimization routines developed at École Polytechnique Fédérale de Lausanne and University of Michigan.

Key mathematical components connect to formulations used by researchers at Brown University, Duke University, Johns Hopkins University, Columbia University, and University of Pennsylvania. The iterative loop uses projections and back-projections in the spirit of techniques from Georgia Institute of Technology, Arizona State University, University of California, Irvine, and University of Sydney. For implementation, engineers reference libraries and toolkits originating from OpenCV, Point Cloud Library, ROS, OpenGL, and software from NVIDIA and AMD.

POSIT’s convergence properties and conditions for unique solution tie to theoretical results studied at Princeton Plasma Physics Laboratory and mathematical departments at Harvard University, University of Chicago, and Rutgers University. The algorithm handles coplanar and non-coplanar model points, drawing comparisons to algorithms analyzed by groups at University of Waterloo, McGill University, Monash University, and University of British Columbia.

Applications

POSIT has been applied in robotics research at Carnegie Robotics, Boston Dynamics, KUKA, and ABB Robotics. Augmented reality systems from Niantic, Magic Leap, HTC, Microsoft HoloLens, and Apple Inc. have employed pose-estimation building blocks related to POSIT in prototype stages. In photogrammetric mapping, agencies such as National Aeronautics and Space Administration and private firms like Trimble Inc., Hexagon AB, Leica Geosystems, and Esri used similar pose recovery methods. Automotive projects at Tesla, Inc., Waymo, Volkswagen AG, and Ford Motor Company referenced pose techniques for sensor fusion research.

Pose estimation for biomechanics and medical imaging appears in studies from Mayo Clinic, Cleveland Clinic, Johns Hopkins Hospital, Massachusetts General Hospital, and academic centers like University College London Hospitals. Research in cultural heritage and archaeology leveraged POSIT-like methods in projects at British Museum, Smithsonian Institution, Louvre, and Getty Conservation Institute. Computer graphics pipelines at Pixar, Walt Disney Animation Studios, Industrial Light & Magic, and DreamWorks Animation have used pose estimators in production toolchains.

Performance and Limitations

POSIT is computationally efficient and was attractive for real-time systems implemented on hardware platforms from Intel Corporation, ARM Holdings, Texas Instruments, and Qualcomm. Benchmarks in literature compare POSIT against techniques from Lucasfilm, Hewlett-Packard, Fujitsu, and academic groups at University of Maryland and City, University of London. Limitations include sensitivity to correspondence errors, occlusions, and dependence on accurate focal length models similar to issues discussed at National Institute of Standards and Technology. Failure cases often examined by researchers at TNO, Fraunhofer Society, and Centre National de la Recherche Scientifique involve near-planar configurations and ambiguous point arrangements.

Robustness improvements derive from integrating RANSAC as developed by researchers at ETH Zurich and Microsoft Research, and from probabilistic filtering approaches from SRI International and Lockheed Martin. Comparisons with bundle adjustment methods popularized by University of Oxford and graph-based SLAM frameworks from Oxford's Visual Geometry Group reveal trade-offs between speed and global optimality. Hardware acceleration via GPUs from NVIDIA and inference offloading to cloud platforms such as Amazon Web Services, Google Cloud Platform, and Microsoft Azure address scaling concerns.

Variants and Extensions

Extensions of POSIT include hybrid formulations that combine it with perspective-n-point solvers studied at University of Bonn, University of Cambridge Machine Intelligence Lab, ETH Zurich Robotics and applied in works from Toyota Research Institute, Honda Research Institute, Intel Labs, and Adobe Research. Robustified variants incorporate statistical estimators championed at Princeton University and University of Texas at Austin. Multi-view and temporal extensions link to visual odometry and SLAM systems developed at Carnegie Mellon University Robotics Institute, Oxford Robotics Institute, MIT CSAIL, and ETH Zurich Autonomous Systems Lab.

Deep learning hybrids integrate convolutional architectures from Heidelberg University, Facebook AI Research, DeepMind, OpenAI, and Google Brain to predict correspondences prior to pose refinement. Industrial adaptations appear in products by Bosch, Siemens, Honeywell, GE Aviation, and Rolls-Royce Holdings. Academic work comparing POSIT-style methods and modern neural pose regressors is common at University of Washington, University of California, San Diego, University of Edinburgh, and University of Hong Kong.

Category:Computer vision algorithms