Toggle search
Toggle menu
5.5K
6.1K
231
53.7K
VR & AR Wiki
Toggle preferences menu
Toggle personal menu
Not logged in
Your IP address will be publicly visible if you make any edits.

Markerless motion capture

vrarwiki.com · VR & AR Wiki
More actions

Markerless motion capture (also written marker-less motion capture, and shortened to markerless mocap) is the recording of human movement from standard video, without the reflective markers, special suits or body-worn sensors used in conventional motion capture. Software identifies the positions and orientations of the body's segments in each image, often with deep learning-based pose estimation algorithms.[1] A 2022 review divided such systems into monocular capture, which uses a single camera, and multi-camera capture, which records video from two or more synchronized cameras.[1]

A 2018 review described automatic markerless systems as an emerging technology, "often driven by the needs of the entertainment industry", that draws on recent trends in computer vision and machine learning.[2] Commercial services turn a single video, recorded on a phone or any other camera, into 3D character animation,[3][4] and multi-camera systems such as CapturyLive track performers in real time and stream the motion into game engines such as Unreal Engine and Unity.[5] With Unreal Engine 5.8, Epic Games added an experimental plugin that captures body motion, or body and facial performance, for its MetaHuman characters from a single camera.[6][7]

The Captury, a maker of markerless systems, says that in virtual reality the technique is used to translate a person's movements into those of an avatar, and that in game development it lets actors be tracked and streamed into engines without suits or markers.[8] In location-based virtual reality, Vicon's markerless system was announced in March 2025 as the tracking technology for The House of Wonders, a six-person experience at Dreamscape Immersive's Geneva site.[9] Validation studies in biomechanics have found markerless measurements of walking close to marker-based ones for some variables, with larger differences at the hip and knee and for rotations about a limb's long axis; the 2022 review concluded that joint center locations and joint angles were not yet accurate enough for clinical use.[10][11][1]

Real-time estimation of a headset user's own body is covered in Body tracking, Full-body tracking and Full-body pose estimation from sparse tracking. Markerless tracking is a different subject: the positional tracking of a headset or other device without fiducial markers.

Reviewed 1 November 2026. Checked the 12 cited DOIs and 5 cited arXiv papers, the study figures from Wade, Colyer, Kanko, Needham, OpenCap and Corazza, and the vendor, documentation and press pages for product names, dates, prices and November 2026 status. About review dates.

How it works

According to Colyer and colleagues' 2018 review, the algorithms that infer body pose from images are usually divided into two families. Generative algorithms use the parameters of a body model to generate a hypothesis, compare it with the image data and refine it iteratively to find the best fit. Discriminative algorithms, often called model-free algorithms, skip that fitting and infer the pose directly from the image data. They learn from a very large database of example data, and compared with generative methods they often process faster, are more robust and depend less on an initial guess, but they can be less precise.[2] Systems based on deep learning describe the posture in each individual image of a video, or of the videos from several cameras.[1]

Keypoint detection

Many systems first locate keypoints, body landmarks such as the shoulders, elbows, hips, knees and ankles, in every frame. Training a detector generally requires a dataset of thousands of manually labeled keypoints, and the 2022 review notes that deep-learning pose estimation "will only ever be as good as the training data used".[1] DeepPose, by Alexander Toshev and Christian Szegedy (CVPR 2014), treated pose estimation as the regression of joint positions by a cascade of deep neural networks.[12]

OpenPose, from the Perceptual Computing Lab at Carnegie Mellon University,[13] uses a representation called Part Affinity Fields to associate detected body parts with the right person in a bottom-up step; its authors report real-time performance regardless of the number of people in the image, and the method placed first in the inaugural COCO 2016 keypoints challenge.[14] The journal version added a combined body and foot keypoint detector.[15] Its developers describe the released library as the first real-time multi-person system to jointly detect body, hand, facial and foot keypoints, 135 in total, on single images.[13]

BlazePose, described by Valentin Bazarevsky and colleagues in a 2020 arXiv paper presented at the CVPR 2020 Workshop on Computer Vision for Augmented and Virtual Reality, is a lightweight convolutional neural network that outputs 33 keypoints for a single person, and its authors report more than 30 frames per second on a Pixel 2 phone.[16] A variant of BlazePose runs in the Pose Landmarker task of Google's MediaPipe, which returns 33 landmarks in image coordinates and in 3D world coordinates from still images, decoded video frames or a live camera stream.[17]

Reconstruction in three dimensions

Multi-camera systems add a reconstruction step that combines the 2D keypoints from several cameras into 3D joint positions, which reduces errors from occluded joints; compared with single-camera setups they cost more and need more space.[1] OpenCap, an open-source system from Stanford University, records with two or more iOS devices, detects keypoints with OpenPose or HRNet and triangulates them. Two long short-term memory networks then predict the positions of 43 anatomical markers from the 20 triangulated keypoints, and joint angles are computed by inverse kinematics with a musculoskeletal model in the OpenSim software.[18]

A single camera is cheaper and needs less space, but self-occlusion, when one body part hides another, is a major source of error: joints go missing for one or more frames, and the opposite limb can be detected instead, for example a right knee labeled as the left.[1] The VNect method of Dushyant Mehta and colleagues, presented at SIGGRAPH 2017, combines a convolutional network that regresses 2D and 3D joint positions with real-time fitting of a kinematic skeleton for temporally stable results. Its authors described it as the first real-time method to capture the full global 3D skeletal pose of a human from a single RGB camera, and as usable in applications such as 3D character control that had previously relied on RGB-D depth cameras.[19]

Body models

Many methods describe the result with a parametric body model rather than a bare skeleton. SMPL (Skinned Multi-Person Linear model), published by Matthew Loper and colleagues at SIGGRAPH Asia 2015, is a skinned, vertex-based model with separate parameters for body shape and pose.[20] Its project page says the model is learned from thousands of 3D body scans and, because it uses blend skinning, is compatible with existing rendering engines. The page provides registered users with example FBX files, code for Python, Maya and Unity, a Blender add-on and an Unreal Engine plugin for research purposes, with commercial use licensed separately.[21]

Human Mesh Recovery (HMR), by Angjoo Kanazawa and colleagues (CVPR 2018), estimates body shape and 3D joint angles directly from a single image. It minimizes the reprojection error of 2D keypoints and adds an adversary, trained on a large database of 3D human meshes, that judges whether the body parameters are realistic; the authors show that the model can be trained with or without paired 2D-to-3D supervision.[22] VIBE, by Muhammed Kocabas, Nikos Athanasiou and Michael J. Black (CVPR 2020), applies the idea to video: an adversarial framework uses the AMASS motion capture dataset to distinguish real human motion from the motion its networks produce, which its authors report yields kinematically plausible sequences.[23]

Physical plausibility and global motion

Methods that reconstruct each frame separately with purely kinematic models can produce physically implausible results. The authors of PhysCap (Soshi Shimada and colleagues, SIGGRAPH Asia 2020) named floor penetration, foot skating, unnatural body leaning and strong shifting in depth among the implausible environment interactions of such methods. PhysCap follows its kinematic estimate with a real-time physics-based pose optimizer that accounts for gravity and collisions, and uses a neural network to detect foot contacts; the system captures motion from a single color camera at 25 frames per second.[24]

WHAM, by Soyong Shin and colleagues (CVPR 2024), targets limitations its authors identified in earlier work: most methods estimate the body in camera coordinates, methods that estimate global position often assume a flat ground plane and produce foot sliding, and the most accurate methods rely on computationally expensive optimization that limits them to offline use. WHAM lifts sequences of 2D keypoints to 3D, uses the camera's angular velocity estimated by a SLAM method to recover the person's global trajectory, and refines that trajectory with a contact-aware method, so that motions such as climbing stairs can be captured.[25] Commercial tools expose comparable corrections as settings. DeepMotion lists physics simulation, foot locking, hand-to-ground contact and motion smoothing for its Animate 3D service,[3] and CG Channel described foot slippage in 2023 as common with video-based motion capture systems, noting that a locomotion filter in Rokoko's Studio software helps to address it.[26]

History

Research on capturing movement without markers took place in both biomechanics and computer vision. The 2018 review by Colyer and colleagues contrasted the two fields: accuracy in computer vision, mainly for entertainment applications, was typically assessed qualitatively and judged on appearance, while biomechanics requires a system to quantify subtle differences in motion.[2]

Many markerless systems have been built on silhouettes, the shape of the person segmented from the background of the image. With enough cameras, silhouettes can be combined into a visual hull, an approximation of the space the person occupies.[2] In 2006 Stefano Corazza and colleagues at Stanford University presented a model-based system that matched a model of the subject to the visual hull with a custom simulated annealing algorithm.[27] The 2018 review credited a later silhouette-based system by Corazza and colleagues, published in 2010 and combining visual hulls, articulated iterative closest point matching and the generation of subject-specific models, with some of the most accurate results reported for automatic markerless body motion capture. It added that such carefully calibrated systems had been limited to laboratory conditions with a large number of cameras, and that the trend had been to move away from silhouettes to improve robustness, reduce ambiguities and reduce the number of cameras.[2] Corazza co-founded the character animation company Mixamo with Nazim Kareemi and served as its CTO. The company's July 2009 launch press release said that Corazza's research at Stanford had "produced the first practical markerless motion capture system based on off-the-shelf video cameras".[28]

Consumer depth cameras offered a second route. The body part recognition algorithm for Kinect, published at CVPR 2011 by Jamie Shotton and colleagues at Microsoft, classifies each pixel of a single depth image into body parts and derives proposals for 3D joint positions from the result; it runs at 200 frames per second on consumer hardware, and the paper received the CVPR Best Paper Award.[29][30][31] The 2018 review found that the Kinect's body tracking, designed for interactive systems, fell well below the accuracy required for precision biomechanics, and listed hardware limits of active depth cameras: short range, frame rates of at most 30 Hz, inoperability in bright sunlight and interference between multiple sensors.[2]

Learning-based methods and models followed: DeepPose in 2014, the SMPL body model in 2015, OpenPose (a 2016 preprint presented at CVPR 2017), the single-camera 3D methods VNect in 2017 and HMR in 2018, VIBE and PhysCap in 2020, and WHAM in 2024, as described above. The 2022 review described OpenPose and DeepLabCut as having become very popular for biomechanical applications.[1] The most recent release of the OpenPose library listed on GitHub, version 1.7.0, dates from 17 November 2020.[32]

The Captury was founded in Saarbrücken in 2013 to market markerless technology developed by researchers around Nils Hasler in Christian Theobalt's group at the Max Planck Institute for Informatics, and was bought by Dari Motion Inc. in 2021.[33] Vicon, a supplier of optical motion capture, launched its own markerless system at the Game Developers Conference in March 2025.[34] Among services that work from a single ordinary video, DeepMotion launched its cloud service Animate 3D on 29 October 2020,[35] Rokoko launched its free browser-based tool Rokoko Vision, the successor to its Rokoko Video system of the previous year, in September 2023,[26] and Move AI released its single-iPhone app Move One in November 2023.[36]

From 2024, Autodesk and Epic Games bought companies and technology in the field. Autodesk announced its purchase of Wonder Dynamics in May 2024 and renamed that company's Wonder Studio as Autodesk Flow Studio in March 2025,[37][38] and in April 2026 it acquired the core technology of RADiCAL.[39] In February 2026 the Max Planck Society announced that Epic Games was acquiring Meshcapade, a spin-off founded in 2018 whose technology captures 3D human motion from standard video and is based on SMPL.[40]

Commercial systems

The table lists systems documented as of November 2026. Capabilities and performance figures are the vendors' own statements unless another source is named.

System Developer Capture setup Processing and output Status
CapturyLive and CapturyStudio The Captury Four to 12 or more cameras for CapturyLive Real-time body tracking, with finger and face tracking offered as add-on modules; streams to Unreal Engine through a Live Link plugin and to Unity and MotionBuilder; exports FBX, BVH, C3D and CSV[5] Available; owned by Dari Motion since 2021[33]
Vicon Markerless Vicon Vanguard markerless cameras Shōgun Markerless and Shōgun Post software, integrated with Vicon's optical pipeline; aimed at VFX, virtual production and previsualization[34] Launched at GDC in March 2025; generally available for VFX from 19 March 2025[34]
Genesis Move AI Six to 12 Z-Cam cameras with genlock and timecode, in volumes of up to 20 by 20 m; one to eight performers[41][42] A desktop application processes the video on a local NVIDIA GPU instead of in the cloud, with live pre-visualization of the motion data during capture; the company says the data is "comparable to optical systems"[42][41] Unveiled in April 2026, when it replaced Move Pro on Move AI's website;[42] presented as the company's flagship system on its products page in November 2026[41]
Move Live Move AI Four to eight machine vision cameras; up to two performers Real time on an on-premises Linux server, with a latency CG Channel gave as around 100 ms; streams to Unreal Engine 5 through a Live Link plugin[43] Launched June 2024 and sold by subscription;[43] not listed on Move AI's products page in November 2026[41]
Move One (App Store name: Move AI 3D Motion) Move AI One iPhone; one performer Cloud processing; FBX and USD export with finger animation but no facial capture at launch[36] Released in November 2023; Gen 2 models (s2 and s2-light) added in version 1.16 (April 2025); version 1.20 of 29 September 2025 was the latest listed in November 2026[36][44]
Animate 3D DeepMotion A single video from any device; up to eight people Cloud service in the browser; exports FBX, BVH, GLB and MP4; face and hand tracking; retargets motion to custom FBX, GLB or VRM characters[3] Launched 29 October 2020; free accounts available[35][3]
Rokoko Vision Rokoko A single video Vision 3.0 runs in the browser-based Rokoko Create with a monocular solver Rokoko says was rebuilt from the ground up; exports FBX or BVH; no finger tracking at launch[45] Available; webcam recording and dual-camera capture removed; free tier of 30 seconds of processing per month[45]
Autodesk Flow Studio (formerly Wonder Studio) Autodesk (from Wonder Dynamics) Live-action video Cloud service that replaces an actor with a CG character and outputs motion capture data, camera tracks and clean plates; exports scenes to Blender, Maya and Unreal Engine, and USD for applications such as Unity[38] Renamed in March 2025, when standard prices were US$29.99 per month for Lite and US$149.99 per month for Pro[38]
Radical Core and Radical Live RADiCAL A single video Radical Core: offline extraction to FBX; Radical Live: cloud streaming to Blender, Maya, Unity or Unreal Engine 5[39] Core technology acquired by Autodesk in April 2026; CG Channel reported that the web portal would close on 6 July 2026 and that existing user data would be deleted after the closure[39]
Plask Plask A single video, including several people Online service using credits; exports FBX, BVH or GLB; Plask Live streams to Blender and Unreal Engine[4] Available[4]
Theia3D Theia Markerless Multi-camera video Deep learning and inverse kinematics; Theia states it tracks 124 keypoints with precision under 1 cm and 3 degrees[46] Intended for research, sports performance and educational use; Theia states it is not cleared as a medical device[46]
MetaHuman Animator Markerless Motion Capture plugin Epic Games A single camera for body, or body and face Processes footage locally on Windows into animation sequences for MetaHuman characters in Unreal Engine[6] Experimental feature introduced in Unreal Engine 5.8; Digital Production reported it as free, including for commercial use[6][7]
Meshcapade Meshcapade Standard video Online platform based on the SMPL body model[40] Acquisition by Epic Games announced in February 2026; the company's website says it is now part of Epic Games and that its online platforms were shut down on 18 April 2026[40][47]

Use in VR game development

For authored character animation, captured motion is exported as skeletal animation and retargeted onto a game character's skeleton. DeepMotion retargets motion automatically onto characters uploaded in FBX, GLB or VRM format.[3] Plask imports avatars in version 0.0 of VRM,[4] a file format based on glTF 2.0 for humanoid 3D avatars in VR applications.[48] Rokoko says the FBX and BVH files exported from its Studio software are compatible with skeletons including Unreal Engine 5's Manny, HIK and Mixamo, and work directly in Blender, Unreal Engine, Unity, Maya and MotionBuilder, among other programs.[45]

Several systems also stream motion into a running engine. CapturyLive streams skeletons to Unreal Engine through its Captury Live Link plugin and supports Unity, MotionBuilder and protocols such as OSC, VRPN, ROS and OSVR.[5] At its launch in 2024, Move Live streamed to Unreal Engine 5 and was aimed at broadcast, film, mixed reality and live events,[43] and Plask Live streams to Blender and Unreal Engine.[4] RADiCAL's cloud service Radical Live streamed animation to Blender, Maya, Unity or Unreal Engine 5; CG Channel noted that its scenes had never been available for download, and the web portal that provided the service was scheduled to close on 6 July 2026.[39]

Vendors and their clients have documented uses in VR. The Captury lists live virtual and location-based entertainment among its applications, saying that its tracking provides interactive experiences "without markers or intrusive equipment".[8] Oxford Metrics, Vicon's parent company, announced on 28 March 2025 that Vicon's markerless system would power The House of Wonders, a six-person location-based VR experience made with the watchmaker Audemars Piguet for Dreamscape's flagship site in Geneva. The announcement said the system allows free movement without markers and with less equipment for users. Sylvain Chagué, co-founder and chief technology officer of Artanim and Dreamscape, said that virtual body ownership and immersion in VR demand both accurate tracking and very low latency, and that the team had dedicated substantial research and development effort to evaluating the computational performance of machine learning-based tracking algorithms and to refining a multi-modal tracking solution that integrates full-body markerless motion capture with VR headset tracking.[9] Dreamscape Immersive was among the beta participants Vicon named at its launch, with Industrial Light & Magic and Gearbox Entertainment.[34]

Accuracy research

Marker-based motion capture has generally served as the reference when markerless accuracy is measured.[1] Quantitative studies come mainly from biomechanics, since computer vision work aimed at entertainment has typically been judged on appearance.[2] None of the studies in the table assessed animation produced for games.

Study Systems compared Participants and task Reported results
Corazza et al., Annals of Biomedical Engineering (2006)[27] Model-based visual hull system with simulated annealing Simulated subject in a virtual environment over a complete gait cycle; also demonstrated on a real subject in a gait laboratory Mean errors of 1.5 degrees (±3.9) for knee flexion, 2.0 degrees (±3.0) for hip flexion, 2.0 degrees (±2.3) for knee adduction and 1.1 degrees (±1.7) for hip adduction
Kanko et al., Journal of Biomechanics (2021)[10][1] Theia3D with eight video cameras, against seven infrared marker-based cameras recording at the same time 30 healthy adults walking on a treadmill Average joint center RMSD below 2.5 cm except at the hip (3.6 cm); segment angle RMSD below 5.5 degrees except for rotations about a segment's long axis. One co-author was affiliated with Theia Markerless Inc.
Needham et al., Scientific Reports (2021)[11] OpenPose, AlphaPose and DeepLabCut on multi-camera 200 Hz video, against marker-based capture 15 healthy participants walking, running and performing counter-movement jumps Systematic hip and knee joint center differences of about 30-50 mm, which the authors consider most likely due to mislabeled ground truth in the training datasets; ankle differences of 1-15 mm depending on the activity
Uhlrich et al., PLOS Computational Biology (2023)[18] OpenCap with two iPhones, against marker-based capture and force plates Ten healthy adults walking, squatting, rising from a chair and performing drop jumps Mean absolute error of 4.5 degrees for joint angles, 6.2% of bodyweight for ground reaction forces and 1.2% of bodyweight times height for joint moments

Needham and colleagues concluded that 3D joint center locations from these pose estimation methods were "not yet consistently comparable to marker-based motion capture".[11] Kanko and colleagues concluded that the system they tested would be a suitable alternative where the practical benefits of markerless data collection are preferred.[10] The OpenCap authors also reported a 100-subject field study in which a clinician estimated musculoskeletal dynamics 25 times faster than with a laboratory-based approach, at less than 1% of the cost.[18]

The 2022 review by Wade and colleagues found that markerless temporospatial measures, such as step length, generally appeared equivalent to marker-based ones, but that joint center locations and joint angles were not yet accurate enough for clinical applications. It gave typical monocular 3D joint center differences of 40-60 mm from reference methods, and summarized a 2019 study in which lower-limb joint angles from an eight-camera Captury system, which the review says identified the person's silhouette rather than using deep learning, differed from a 16-camera marker-based system by 4-20 degrees. It also pointed out that marker-based capture is itself not a gold standard, so the true accuracy of markerless methods is unknown without comparison against a method such as biplanar videoradiography.[1] The review attributed Theia3D's comparatively strong results to the company having labeled its own biomechanical dataset of 51 keypoints, while calling the proprietary software "somewhat of a black box" because the keypoints in use are not known.[1] As of November 2026, Theia's website gives a figure of 124 keypoints.[46]

Limitations

Capturing 3D motion from one camera is, in the words of the PhysCap authors, "a very challenging and severely ill-posed problem", and physically implausible poses from existing methods are "problematic for any use case in computer graphics".[24] Methods that estimate global motion on an assumed flat ground plane can produce foot sliding.[25]

Training data limits accuracy. Open-source pose estimation datasets were labeled by people recruited from the general population who likely lacked anatomical knowledge, and the 2022 review found that these datasets had not been labeled with the accuracy biomechanics requires.[1] Rotations about the long axis of a limb showed larger differences: in a Theia3D study the review cites, they differed from marker-based values by 6.9-13.2 degrees, against 2.6-11 degrees for flexion, extension, abduction and adduction. The review notes that detecting rotation about the longitudinal axis requires three points on a segment.[1] Validation studies have generally been run with participants in tight-fitting clothing, because marker-based reference systems need it; loose clothing may change shape during movement, and its effect on markerless accuracy had not been established.[1] Hands and faces are not always covered: Rokoko Vision 3.0 launched without finger tracking,[45] Plask's finger capture works only with its V1 model,[4] and Move One captured fingers but not the face at its release.[36] Facial capture is covered in Facial motion capture.

Licensing and data handling vary. OpenPose is free for non-commercial use, with commercial licensing handled through Carnegie Mellon University;[13] the VNect project page restricts its code to non-commercial research;[49] SMPL is distributed for research purposes, with commercial use licensed separately;[21] and the WHAM authors announced their code for research purposes.[25] The 2023 OpenCap paper states that the cloud-deployed academic version, hosted at Stanford University, will remain freely available to the academic research community for the foreseeable future, and that Model Health, Inc., a company co-founded by two of its authors, supports commercial use.[18] For cloud services, CG Channel reported in March 2025 that Flow Studio's terms of service gave Autodesk a non-exclusive license to use any content created on the platform to develop its AI models.[38] With Genesis, Move AI moved processing from the cloud to local GPUs, which 80.lv reported as a data-security measure,[42] and Epic's MetaHuman plugin processes footage locally.[6] Hosted services can also close: CG Channel was told that all existing RADiCAL user data would be deleted after the company's web portal closed on 6 July 2026,[39] and Meshcapade shut down its online platforms on 18 April 2026, having become part of Epic Games.[47]

See also

References

  1. ↑ 1.00 1.01 1.02 1.03 1.04 1.05 1.06 1.07 1.08 1.09 1.10 1.11 1.12 1.13 1.14 L. Wade, L. Needham, P. McGuigan, J. Bilzon (2022). "Applications and limitations of current markerless motion capture methods for clinical gait biomechanics". PeerJ, vol. 10, e12995. https://doi.org/10.7717/peerj.12995. Retrieved 2026-11-01.
  2. ↑ 2.0 2.1 2.2 2.3 2.4 2.5 2.6 S. L. Colyer, M. Evans, D. P. Cosker, A. I. T. Salo (2018). "A Review of the Evolution of Vision-Based Motion Analysis and the Integration of Advanced Computer Vision Methods Towards Developing a Markerless System". Sports Medicine - Open, vol. 4, article 24. https://doi.org/10.1186/s40798-018-0139-y. Retrieved 2026-11-01.
  3. ↑ 3.0 3.1 3.2 3.3 3.4 "Animate 3D". DeepMotion. https://www.deepmotion.com/animate-3d. Retrieved 2026-11-01.
  4. ↑ 4.0 4.1 4.2 4.3 4.4 4.5 "Plask - AI Motion Capture from Video". Plask. https://plask.ai/. Retrieved 2026-11-01.
  5. ↑ 5.0 5.1 5.2 "Real-Time Processing". The Captury. https://captury.com/real-time-processing/. Retrieved 2026-11-01.
  6. ↑ 6.0 6.1 6.2 6.3 "MetaHuman Animation from Mono Video Capture in Unreal Engine". MetaHuman Documentation. Epic Games. https://dev.epicgames.com/documentation/metahuman/metahuman-animation-from-mono-video-capture-in-unreal-engine. Retrieved 2026-11-01.
  7. ↑ 7.0 7.1 Bela Beier (2026-06-23). "MetaHuman gets markerless mocap". Digital Production. https://digitalproduction.com/2026/06/23/metahuman-gets-markerless-mocap/. Retrieved 2026-11-01.
  8. ↑ 8.0 8.1 "Captury Applications". The Captury. https://captury.com/application/. Retrieved 2026-11-01.
  9. ↑ 9.0 9.1 "Vicon's new markerless system enabling Dreamscape Immersive's latest VR experience". Oxford Metrics. 2025-03-28. https://oxfordmetrics.com/news/2025-03-28/vicons-new-markerless-system-enabling-dreamscapes-latest-vr-experience. Retrieved 2026-11-01.
  10. ↑ 10.0 10.1 10.2 R. M. Kanko, E. K. Laende, E. M. Davis, W. S. Selbie, K. J. Deluzio (2021). "Concurrent assessment of gait kinematics using marker-based and markerless motion capture". Journal of Biomechanics, vol. 127, 110665. https://doi.org/10.1016/j.jbiomech.2021.110665. Retrieved 2026-11-01.
  11. ↑ 11.0 11.1 11.2 L. Needham, M. Evans, D. P. Cosker, L. Wade, P. M. McGuigan, J. L. Bilzon, S. L. Colyer (2021). "The accuracy of several pose estimation methods for 3D joint centre localisation". Scientific Reports, vol. 11, 20673. https://doi.org/10.1038/s41598-021-00212-x. Retrieved 2026-11-01.
  12. ↑ A. Toshev, C. Szegedy (2014). "DeepPose: Human Pose Estimation via Deep Neural Networks". IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2014, pp. 1653-1660. https://doi.org/10.1109/CVPR.2014.214. Retrieved 2026-11-01.
  13. ↑ 13.0 13.1 13.2 "CMU-Perceptual-Computing-Lab/openpose". GitHub. https://github.com/CMU-Perceptual-Computing-Lab/openpose. Retrieved 2026-11-01.
  14. ↑ Z. Cao, T. Simon, S.-E. Wei, Y. Sheikh (2017). "Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields". IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2017; arXiv:1611.08050. https://arxiv.org/abs/1611.08050. Retrieved 2026-11-01.
  15. ↑ Z. Cao, G. Hidalgo, T. Simon, S.-E. Wei, Y. Sheikh (2021). "OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields". IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 1, pp. 172-186. https://doi.org/10.1109/TPAMI.2019.2929257. Retrieved 2026-11-01.
  16. ↑ V. Bazarevsky, I. Grishchenko, K. Raveendran, T. Zhu, F. Zhang, M. Grundmann (2020-06-17). "BlazePose: On-device Real-time Body Pose tracking". arXiv preprint; CVPR 2020 Workshop on Computer Vision for Augmented and Virtual Reality. https://arxiv.org/abs/2006.10204. Retrieved 2026-11-01.
  17. ↑ "Pose landmark detection guide". Google AI Edge. Google. 2026-10-09. https://developers.google.com/edge/mediapipe/solutions/vision/pose_landmarker. Retrieved 2026-11-01.
  18. ↑ 18.0 18.1 18.2 18.3 S. D. Uhlrich, A. Falisse, Ł. Kidziński, J. Muccini, M. Ko, A. S. Chaudhari, J. L. Hicks, S. L. Delp (2023-10-19). "OpenCap: Human movement dynamics from smartphone videos". PLOS Computational Biology, vol. 19, no. 10, e1011462. https://doi.org/10.1371/journal.pcbi.1011462. Retrieved 2026-11-01.
  19. ↑ D. Mehta, S. Sridhar, O. Sotnychenko, H. Rhodin, M. Shafiei, H.-P. Seidel, W. Xu, D. Casas, C. Theobalt (2017). "VNect: Real-time 3D Human Pose Estimation with a Single RGB Camera". ACM Transactions on Graphics, vol. 36, no. 4 (SIGGRAPH 2017). https://doi.org/10.1145/3072959.3073596. Retrieved 2026-11-01.
  20. ↑ M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, M. J. Black (2015). "SMPL: A Skinned Multi-Person Linear Model". ACM Transactions on Graphics, vol. 34, no. 6 (SIGGRAPH Asia 2015). https://doi.org/10.1145/2816795.2818013. Retrieved 2026-11-01.
  21. ↑ 21.0 21.1 "SMPL". Max Planck Institute for Intelligent Systems. https://smpl.is.tue.mpg.de/. Retrieved 2026-11-01.
  22. ↑ A. Kanazawa, M. J. Black, D. W. Jacobs, J. Malik (2018). "End-to-end Recovery of Human Shape and Pose". IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2018; arXiv:1712.06584. https://arxiv.org/abs/1712.06584. Retrieved 2026-11-01.
  23. ↑ M. Kocabas, N. Athanasiou, M. J. Black (2020). "VIBE: Video Inference for Human Body Pose and Shape Estimation". IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2020; arXiv:1912.05656. https://arxiv.org/abs/1912.05656. Retrieved 2026-11-01.
  24. ↑ 24.0 24.1 S. Shimada, V. Golyanik, W. Xu, C. Theobalt (2020). "PhysCap: Physically Plausible Monocular 3D Motion Capture in Real Time". ACM Transactions on Graphics, vol. 39, no. 6 (SIGGRAPH Asia 2020). https://doi.org/10.1145/3414685.3417877. Retrieved 2026-11-01.
  25. ↑ 25.0 25.1 25.2 S. Shin, J. Kim, E. Halilaj, M. J. Black (2024). "WHAM: Reconstructing World-grounded Humans with Accurate 3D Motion". IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2024, pp. 2070-2080; arXiv:2312.07531. https://arxiv.org/abs/2312.07531. Retrieved 2026-11-01.
  26. ↑ 26.0 26.1 Jim Thacker (2023-09-04). "Check out free browser-based AI mocap tool Rokoko Vision". CG Channel. https://www.cgchannel.com/2023/09/check-out-free-browser-based-ai-mocap-tool-rokoko-vision/. Retrieved 2026-11-01.
  27. ↑ 27.0 27.1 S. Corazza, L. Mündermann, A. M. Chaudhari, T. Demattio, C. Cobelli, T. P. Andriacchi (2006). "A Markerless Motion Capture System to Study Musculoskeletal Biomechanics: Visual Hull and Simulated Annealing Approach". Annals of Biomedical Engineering, vol. 34, no. 6, pp. 1019-1029. https://doi.org/10.1007/s10439-006-9122-8. Retrieved 2026-11-01.
  28. ↑ "3D Pioneers Unveil New Company - Mixamo". Animation World Network. 2009-07-27. https://www.awn.com/news/3d-pioneers-unveil-new-company-mixamo. Retrieved 2026-11-01.
  29. ↑ J. Shotton, A. Fitzgibbon, M. Cook, T. Sharp, M. Finocchio, R. Moore, A. Kipman, A. Blake (2011). "Real-time human pose recognition in parts from single depth images". IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2011, pp. 1297-1304. https://doi.org/10.1109/CVPR.2011.5995316. Retrieved 2026-11-01.
  30. ↑ "Real-Time Human Pose Recognition in Parts from a Single Depth Image". Microsoft Research. Microsoft. 2011-06. https://www.microsoft.com/en-us/research/publication/real-time-human-pose-recognition-in-parts-from-a-single-depth-image/. Retrieved 2026-11-01.
  31. ↑ Rob Knies (2011-09-26). "Kinect Body Tracking Reaps Renown". Microsoft Research Blog. Microsoft. https://www.microsoft.com/en-us/research/blog/kinect-body-tracking-reaps-renown/. Retrieved 2026-11-01.
  32. ↑ "Releases - CMU-Perceptual-Computing-Lab/openpose". GitHub. https://github.com/CMU-Perceptual-Computing-Lab/openpose/releases. Retrieved 2026-11-01.
  33. ↑ 33.0 33.1 "Dari Motion buys Max Planck spin-off The Captury". Max Planck Institute for Informatics. 2021. https://mpi-inf.mpg.de/news/press-releases/2021/dari-motion-buys-max-planck-spin-off-the-captury. Retrieved 2026-11-01.
  34. ↑ 34.0 34.1 34.2 34.3 "Vicon Launches Markerless Motion Capture at GDC". Vicon. 2025-03-11. https://www.vicon.com/resources/press/markerlesslaunch/. Retrieved 2026-11-01.
  35. ↑ 35.0 35.1 "Motion Technology Startup DeepMotion Makes Motion Capture Accessible to Everyone with Launch of Animate 3D". DeepMotion. 2020-10-29. https://deepmotion.com/post/motion-technology-startup-deepmotion-makes-motion-capture-accessible-to-everyone-with-launch-of-animate-3d-2. Retrieved 2026-11-01.
  36. ↑ 36.0 36.1 36.2 36.3 Jim Thacker (2023-11-07). "Move AI releases Move One". CG Channel. https://www.cgchannel.com/2023/11/move-ai-releases-move-one. Retrieved 2026-11-01.
  37. ↑ "Autodesk Acquires Wonder Dynamics". Animation World Network. 2024-05-21. https://www.awn.com/news/autodesk-acquires-wonder-dynamics. Retrieved 2026-11-01.
  38. ↑ 38.0 38.1 38.2 38.3 Jim Thacker (2025-03-14). "Wonder Studio becomes Autodesk Flow Studio". CG Channel. https://www.cgchannel.com/2025/03/wonder-studio-becomes-autodesk-flow-studio. Retrieved 2026-11-01.
  39. ↑ 39.0 39.1 39.2 39.3 39.4 Jim Thacker (2026-04-09). "Autodesk acquires core tech of AI motion-capture firm Radical". CG Channel. https://www.cgchannel.com/?p=175200. Retrieved 2026-11-01.
  40. ↑ 40.0 40.1 40.2 "Max Planck spin-off draws Epic Games to Tübingen". Max Planck Society. 2026-02-18. https://www.mpg.de/26082348/max-planck-spin-off-draws-epic-games-to-tubingen. Retrieved 2026-11-01.
  41. ↑ 41.0 41.1 41.2 41.3 "Move.ai Products - Leading Markerless Motion Capture Solutions". Move AI. https://www.move.ai/products. Retrieved 2026-11-01.
  42. ↑ 42.0 42.1 42.2 42.3 Gloria Levine (2026-04-09). "Move Reveals Its New Markerless Motion Capture System Genesis". 80.lv. https://80.lv/articles/move-s-new-markerless-motion-capture-system-genesis-promises-cinema-quality-imaging. Retrieved 2026-11-01.
  43. ↑ 43.0 43.1 43.2 Jim Thacker (2024-06-27). "Move AI launches AI-based real-time mocap solution Move Live". CG Channel. https://www.cgchannel.com/?p=159104. Retrieved 2026-11-01.
  44. ↑ "Move AI 3D Motion". App Store. Apple. https://apps.apple.com/us/app/move-one/id6448635527. Retrieved 2026-11-01.
  45. ↑ 45.0 45.1 45.2 45.3 "Rokoko Vision: AI motion capture from video, start for free". Rokoko. https://www.rokoko.com/products/vision. Retrieved 2026-11-01.
  46. ↑ 46.0 46.1 46.2 "Theia - Markerless Motion Capture". Theia Markerless. https://www.theiamarkerless.com/. Retrieved 2026-11-01.
  47. ↑ 47.0 47.1 "Meshcapade". Meshcapade. https://meshcapade.com/. Retrieved 2026-11-01.
  48. ↑ "VRM Consortium". VRM Consortium. https://vrm-consortium.org/en/. Retrieved 2026-11-01.
  49. ↑ "VNect: Real-time 3D Human Pose Estimation with a Single RGB Camera, SIGGRAPH 2017". Max Planck Institute for Informatics. https://vcai.mpi-inf.mpg.de/projects/VNect/. Retrieved 2026-11-01.
Retrieved from "https://vrarwiki.com/index.php?title=Markerless_motion_capture&oldid=53291"