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.

Full-body pose estimation from sparse tracking

vrarwiki.com · VR & AR Wiki
More actions

Full-body pose estimation from sparse tracking is the reconstruction of a person's whole-body posture from the few body points that a virtual reality (VR) or mixed reality system measures directly. In a typical consumer setup these are the positions and orientations of the headset and of the two hand controllers or tracked hands, a configuration known as three-point tracking; some systems add a few extra trackers or inertial measurement units (IMUs).[1][2] In the three-point case no sensor observes the hips, knees or feet, so the lower body has to be inferred from upper-body signals and is the hardest part to reconstruct.[3][4] Machine learning methods such as AvatarPoser and AGRoL are trained on motion capture recordings and output the joint rotations of a parametric body model fast enough to animate an avatar in real time.[5][3]

Paper titles also describe the task as "full-body tracking from sparse observations" (AvatarJLM) or "avatar generation from sparse observations" (FLAG and SAGE).[6][7][4] Published methods include Transformer regression networks such as AvatarPoser (ETH Zurich, 2022), diffusion models such as AGRoL (Meta, 2023) and physics-based controllers such as QuestSim (Meta, 2022).[5][3][8] In products, Meta's Movement SDK infers a body pose from headset, controller and hand movement on Quest headsets, and its Generative Legs feature, released to developers in December 2023, generates leg motion without leg sensors.[9][10]

The technique is one form of body tracking. Inverse kinematics (IK) solvers pose the untracked joints geometrically from the tracked end points, full-body tracking adds worn trackers to measure more of the body, and markerless motion capture reconstructs body motion from camera images.

Reviewed 1 November 2026. Cited papers' metadata and results against their full text or abstracts, and product and license statements against the cited Meta, press, GitHub, Sony and AMASS pages. About review dates.

Problem definition

A headset and two controllers each report a six-degree-of-freedom pose.[1] AvatarPoser, for example, gives its network the position, rotation, linear velocity and angular velocity of each of the three devices, 54 values per frame.[5] Other methods use more signals: LoBSTr also takes the tracked pelvis,[11] SparsePoser works from six tracking devices,[12] DivaTrack adds linear accelerations from IMUs on the headset and the wrists,[1] and HMD-Poser accepts optional IMUs on the lower legs and the pelvis.[13] QuestSim can also drive its avatar from the headset pose alone.[8]

AvatarPoser, AGRoL and AvatarJLM output joint rotations of SMPL (Skinned Multi-Person Linear model), a body model learned from a relatively large number of aligned 3D meshes of different people in different poses.[14][5][3][6] AvatarPoser predicts the rotations of the first 22 joints of the SMPL kinematic tree and uses the rotation at the pelvis, the root of the model, as the global orientation of the body.[5] Meta's Movement SDK instead uses a proprietary skeleton of 84 bones; its full-body joint set adds 14 lower-body joints to 70 upper-body joints.[15]

The SAGE authors point out that in SMPL's kinematic tree the upper and lower body share only one common ancestor node, which they see as an opportunity to reconstruct the two halves separately.[4] DivaTrack's authors report that earlier methods lost quality in less common activities such as running backward and carrying objects, and in cases where the lower-body pose is ambiguous, such as kicking, lunging, crouching or looking around while walking.[1]

Real-time operation adds further constraints. An online system can use only current and past input, while reading future frames adds context at the cost of delay: QuestSim reads six future observations, which its authors say introduces 160 ms of latency.[8] In the SAGE authors' comparison, AGRoL run offline, processing a whole input sequence in one pass, had a mean per joint velocity error of 18.59 cm/s, the figure published with AGRoL, but 79.12 cm/s when run online.[4][3] The EgoPoser authors found that AvatarPoser, AGRoL and AvatarJLM, which take global input representations, lost accuracy when the test sequences were shifted by up to 50 m to simulate a user at a different location, while CoolMoves and Final IK, which use local representations, did not.[2] With camera-based hand tracking the hands are observed only inside the headset cameras' field of view, a condition that HMD-NeMo (2023) and EgoPoser (2024) model explicitly.[16][2]

Approaches

The AvatarPoser authors used RootMotion's Final IK solver, run in Unity, as a conventional baseline; it gave the largest rotation and position errors in their comparison, which they attribute to its optimizing the poses of the end effectors without considering the smoothness of the other body joints.[5] An earlier data-driven system, CoolMoves (2021) by Karan Ahuja, Mar Gonzalez-Franco and colleagues, synthesizes accentuated full-body avatar motion in real time from headset and hand positions by matching the user's motion to similar motions in a motion capture database and interpolating between them; its authors report that it works well even with very sparse motion databases, for example with only three to five motions per action.[17] In their reviews of earlier work, the authors of HMD-Poser and SAGE divide learned methods into regression models and generative models, and the HMD-Poser authors list physics-simulator-based methods as a separate category.[13][4]

Regression models

Regression models map the recent input directly to a single pose estimate. LoBSTr (KAIST, 2021) is a recurrent neural network built from gated recurrent units that predicts the lower-body pose and foot contact states from the velocities of the head, hand and pelvis signals; the predicted contacts are used to post-process the lower body with inverse kinematics, removing foot-skating and floating artifacts.[11]

AvatarPoser, presented at ECCV 2022 by Jiaxi Jiang and colleagues at ETH Zurich with two co-authors from Reality Labs at Meta, combines a Transformer encoder with a forward-kinematics module and an inverse kinematics optimization that refines the arm joints so that the hands match the tracked input. The authors describe it as the first learning-based method to predict full-body poses in world coordinates from head and hand input alone, and report inference rates of up to 662 frames per second. In their ablation on the HDM05 test set, removing the IK step barely changed the full-body position error (8.05 cm with it, 8.07 cm without) but raised the hand position error from 1.86 to 3.17 cm. They also tested it on recordings from an HTC Vive Pro headset and two Vive controllers.[5]

AvatarJLM (ByteDance, ICCV 2023) first builds features for each joint and then processes them as tokens in alternating spatial and temporal Transformer blocks; it is trained with loss terms for hand alignment, motion smoothness and physical plausibility.[6] HMD-Poser (CVPR 2024), by researchers at ByteDance's PICO, handles headset-and-controller input alone or with two or three extra IMUs in a single network, filling the inputs of missing sensors with zeros. It combines long short-term memory layers with a Transformer encoder, estimates the user's body shape online and, according to its authors, runs at 90 Hz on a PICO 4 headset.[13] EgoPoser (ETH Zurich, ECCV 2024) simulates the loss of hand tracking whenever the hands leave a modeled headset field of view, decomposes global motion so that it predicts the full-body pose independently of global position, captures a longer motion history with an efficient SlowFast module and estimates body shape; its authors report an inference speed of over 600 frames per second.[2] EnvPoser (CVPR 2025) regresses the motion together with an uncertainty estimate for each joint, which its authors use to model the several poses that can fit the same input, and then refines the estimate with semantic and geometric constraints from a pre-scanned environment.[18]

Generative models

Generative models learn a distribution of plausible poses for a given input instead of a single answer. VAE-HMD (ICCV 2021) uses variational autoencoders to generate articulated poses of a human skeleton from noisy streams of head and hand pose.[19] FLAG (Microsoft, CVPR 2022) is a flow-based generative model that learns a conditional distribution of 3D body pose and generates a plausible pose with uncertainty estimates for the joints.[7] AGRoL ("Avatars Grow Legs", Meta AI, CVPR 2023) is a conditional diffusion model with a multilayer perceptron backbone. It is trained with 1,000 diffusion steps but sampled with five DDIM steps at inference, producing 196 frames in 35 ms on an NVIDIA V100 GPU. When its authors randomly masked 10% of the input frames during inference to simulate tracking loss, AGRoL degraded less than the earlier methods they tested.[3] BoDiffusion (ICCV 2023 workshops) also generates motion sequences by diffusion, with a scheme that conditions the generation in time and space on the sparse tracking input.[20]

SAGE (CVPR 2024) splits generation into two stages: one latent diffusion model reconstructs the upper body, and a second reconstructs the lower body conditioned on the sparse input and the upper-body latent, both working in latent spaces learned by VQ-VAE models; its authors report 0.74 ms per frame on an NVIDIA RTX 3090 GPU.[4] DivaTrack (KAIST and Meta Reality Labs, Eurographics 2024) pairs a regression model for the upper body and foot contacts with a conditional variational autoencoder for the lower body, blends predictions computed in two complementary reference frames, and infers the user's body proportions from a one-time calibration with six poses.[1]

Physics-based tracking

QuestSim (Meta, presented at SIGGRAPH Asia 2022) trains a control policy with reinforcement learning that outputs joint torques for a physics-simulated character rather than poses. The policy was trained on 8 hours of in-house motion capture from 172 subjects and runs at 36 frames per second on headset data supplied at 72 frames per second. Because a physics simulator generates the motion, the authors note that it does not violate physical laws through artifacts such as foot skating or floor penetration; the character can, however, fall when imitating a dynamic motion it has not learned, such as break-dancing or jumping.[8] QuestEnvSim (SIGGRAPH 2023) adds observations of the environment, so that the simulated avatar can sit on chairs, a couch and boxes, step over boxes, rock a chair and turn an office chair.[21]

Datasets and evaluation

AvatarPoser and AGRoL are trained and evaluated on AMASS (Archive of Motion Capture as Surface Shapes), which unifies 15 optical marker-based motion capture datasets as SMPL bodies fitted with a method called MoSh++, with more than 40 hours of motion from over 300 subjects.[22][5][3] The headset and hand inputs are simulated from the captured bodies; the authors of the Nymeria dataset note that sensor inputs simulated from motion datasets lack the noise characteristics of real data.[23] AMASS is distributed under a license that permits use only for non-commercial scientific research, education or artistic projects and prohibits commercial use, including incorporation in a commercial product.[24]

AGRoL, AvatarJLM and SAGE all report results on the same two AMASS evaluation protocols. The first, published with AvatarPoser, uses the CMU, BMLrub and HDM05 subsets, split at random into 90% training and 10% test data.[5] The second, larger protocol, which AvatarJLM's authors describe as following FLAG, trains on 12 AMASS subsets and tests on the Transition and HumanEva subsets.[3][6][4] AvatarJLM's authors also recorded a PICO 4 headset and controllers together with a synchronized OptiTrack motion capture system;[6] HMD-Poser's real dataset has 74 free-dancing motions from eight subjects wearing a PICO 4 and two PICO motion trackers on the lower legs;[13] and DivaTrack's 16.5-hour dataset covers 22 subjects with different body proportions, recorded with an Xsens system.[1] Nymeria, from Meta Reality Labs Research and presented at ECCV 2024, contains 300 hours of daily activities from 264 participants in 50 locations, recorded with Xsens motion capture suits, Project Aria glasses and two wristbands.[23]

Reported metrics include mean per joint rotation error (MPJRE, in degrees), mean per joint position error (MPJPE, in cm) and mean per joint velocity error (MPJVE, in cm/s).[5] AGRoL, AvatarJLM and SAGE also report jitter, the mean jerk (rate of change of acceleration) of the body joints, as a measure of smoothness, and separate position errors for regions such as the hands, the upper body and the lower body.[3][6][4] AvatarJLM adds ground-contact and foot-skating metrics, and EgoPoser measures ground penetration and foot floating.[6][2]

The table lists results on the first protocol with headset and hand input only. The AvatarPoser authors ran Final IK in Unity and reproduced CoolMoves, LoBSTr and VAE-HMD because those methods had no public source code, adapting the position-based CoolMoves to predict rotations; for a fair comparison, they trained all methods on the same training and test data.[5]

Method MPJRE (degrees) MPJPE (cm) MPJVE (cm/s) Reported in
Final IK 16.77 18.09 59.24 AvatarPoser paper[5]
CoolMoves 5.20 7.83 100.54 AvatarPoser paper[5]
LoBSTr 10.69 9.02 44.97 AvatarPoser paper[5]
VAE-HMD 4.11 6.83 37.99 AvatarPoser paper[5]
AvatarPoser 3.21 4.18 29.40 AvatarPoser paper[5]
AGRoL (whole sequence) 2.66 3.71 18.59 AGRoL paper[3]
AGRoL (online) 2.96 4.26 79.12 SAGE paper[4]
AvatarJLM 2.90 3.35 20.79 AvatarJLM paper[6]
SAGE 2.53 3.28 20.62 SAGE paper[4]

In AGRoL's own evaluation, the lower-body position error (6.84 cm) was more than four times the upper-body error (1.55 cm).[3] QuestSim reports an MPJPE of 3.7 cm for headset and controller input synthesized from a different dataset, LaFAN, so its figures are not directly comparable.[8]

Products and applications

The avatar system that Facebook launched on Quest in April 2021 used what the company called "machine-learning based motion prediction technology": a model, trained on what Facebook had learned from studying physical movement, that predicts shoulder and elbow positions from the positions of the headset and controllers. Outside the avatar editor, the avatars were shown from the waist up.[25]

Meta's Movement SDK documentation states that the Body Tracking API "uses hand, controller, and headset movements to infer the body poses of the user" and that body tracking is supported by all the Quest headsets it lists: Quest 2, Quest Pro, Quest 3 and Quest 3S.[9] In September 2023 Meta announced Inside-Out Body Tracking (IOBT), which uses cameras on the headset to track the upper body, including elbow and wrist movement and leaning from side to side, and Generative Legs, which uses information about the upper body to generate what the lower body might be doing. Meta said Generative Legs would let developers deliver "much more natural movements compared to IK-based approaches" and could be applied on Quest 3, Quest Pro and Quest 2.[26] On 20 December 2023 Meta announced that both features were available for public release as part of v60, with IOBT exclusive to Quest 3 and Generative Legs compatible with all Meta Quest devices.[10] UploadVR described Generative Legs as an estimator rather than tracking; according to UploadVR it can detect jumping and crouching, and it works on Quest Pro and Quest 2 from the head and hands alone.[27] MIXED reported that the leg estimate is more accurate when combined with IOBT data.[28]

Meta's Dodge Arcade demo, which uses both features, was available on App Lab in December 2023,[26][28] and UploadVR reported that the combination, called Full Body Synthesis, would come to Supernatural, Swordsman VR and Drunkn Bar Fight.[27] In March 2024 Virtual Desktop began emulating worn Vive Trackers for SteamVR body tracking: according to UploadVR, it drives the emulated upper-body trackers with IOBT on Quest 3, or with inverse kinematics on older Quest headsets, and the lower-body trackers with Generative Legs. UploadVR noted that Vive Trackers are mainly used by VRChat users.[29] Version 83.0.0 of Meta's Unity-Movement package, released on 16 December 2025, added an AI Motion Synthesizer that, according to the release notes, "uses AI to generate natural, full-body character motion from sparse input signals".[30] Meta's documentation says it blends input from the Quest controllers with body tracking data, removes the need for separate walking, running and turning animations, and lets seated users experience full body tracking as if they were standing; it is available only on Quest headsets or on Windows PCs connected to them.[15]

Other products add a few body-worn sensors. UploadVR reported in August 2024 that each of PICO's Motion Trackers, which add leg tracking to the Pico 4 Ultra, Pico 4 and Pico Neo 3, contains an IMU and 12 infrared LEDs. The headset tracks the LEDs for six-degree-of-freedom positional tracking while a tracker is in view of its cameras; otherwise the IMU data is fed into a skeletal model to produce "plausible (but imperfect)" estimated leg poses. ByteDance claimed an average position error of 5 cm and a latency under 20 ms.[31] Sony's mocopi attaches six sensors with accelerometers and gyro sensors to the head, waist, hands and legs. According to Sony, an AI model interpolates joints that carry no sensor, such as the elbows and knees, whose positions cannot be determined uniquely by a simple geometric calculation, and also corrects the position errors that accumulate when positions are calculated by integrating acceleration.[32]

Limitations

Leg poses remain estimates. Meta says that because Generative Legs are generated rather than tracked, some movements, such as crossing the legs, may not be captured,[26] and UploadVR noted that it does not pick up movements such as raising the knees.[27] The AGRoL authors found that their model struggled with irregular poses that are poorly represented in the training data and with motions in which the lower body is only weakly correlated with the upper body, such as break dancing.[3] AvatarJLM's authors report that their model cannot predict complex lower-body motions such as ballet, may collapse on rare motions such as falling down, and on real recordings has difficulty when the tracking signals vary little while the lower body moves a lot.[6] The SAGE authors report that their model, like the best earlier method they compared, has difficulty with movements induced by external forces and with unconventional poses.[4]

The AGRoL authors call occasional floor penetration artifacts a limitation of their approach and of related ones, and LoBSTr predicts foot contacts specifically to remove foot-skating and floating artifacts.[3][11] Physics-based tracking avoids such artifacts but fails in other ways: besides falling on unfamiliar dynamic motions, QuestSim produces a natural and physically valid lower-body pose that might not match the user's when upper-body and lower-body motion are uncorrelated, and its authors say the simulated avatar can still look stiff and unnatural.[8]

The authors of AvatarJLM, HMD-Poser and Nymeria each evaluated real device recordings and found larger errors than on AMASS or on simulated input. AvatarJLM had an MPJPE of 4.93 cm on the larger AMASS protocol but 9.52 cm on real PICO 4 recordings, while AvatarPoser, retrained by the AvatarJLM authors with its public source code, had 6.44 cm and 11.22 cm.[6] HMD-Poser, running on a PICO 4 with controllers and two lower-leg IMUs, had an MPJPE of 6.55 cm on its real recordings, against 4.75 cm with synthetic input generated from the ground-truth motion; its hand error rose from 2.09 to 8.10 cm, which the authors attribute to the hand's connection to the controller not always being rigid.[13] On Nymeria, AvatarPoser had an MPJPE of 7.97 cm with real device poses and 7.31 cm with simulated inputs, a gap the authors call small; both exceeded the 4.20 cm it reached on AMASS, mainly through worse lower-body tracking, which the authors hypothesize comes from harder motions such as hiking on uneven terrain, taking stairs and playing sports.[23]

Hand visibility is a further limit for camera-tracked hands. In EgoPoser's tests on AMASS, a model trained with fully visible hands had an MPJPE of 24.75 cm when hand tracking was lost outside a simulated 180-degree field of view and 41.24 cm with a 90-degree field of view (the authors give 120 degrees for Quest 2 and 90 degrees for HoloLens 2); with EgoPoser's field-of-view modeling, the errors were 5.31 cm and 6.60 cm.[2]

Body size matters as well. In EgoPoser's AMASS tests on more than 175 subjects between 145 and 207 cm tall, the model trained with the mean body shape had an MPJPE of 6.36 cm, and estimating each user's body shape lowered it to 4.79 cm.[2] Meta's documentation warns that body tracking might scale a mesh to fit different user heights, which can lead to visual artifacts such as the hand meshes disconnecting from the forearm meshes. It says the automatic height calibration runs within the first 10 seconds of the tracking service being created, ideally with the user standing, and that if IOBT is requested after the system is already highly utilized, body tracking starts in a low-fidelity mode without IOBT, losing features such as elbow tracking.[15]

See also

References

  1. ↑ 1.0 1.1 1.2 1.3 1.4 1.5 Dongseok Yang, Jiho Kang, Lingni Ma, Joseph Greer, Yuting Ye, Sung-Hee Lee (2024). "DivaTrack: Diverse Bodies and Motions from Acceleration-Enhanced Three-Point Trackers". Computer Graphics Forum, vol. 43, no. 2 (Eurographics 2024). https://doi.org/10.1111/cgf.15057. Retrieved 2026-11-01.
  2. ↑ 2.0 2.1 2.2 2.3 2.4 2.5 2.6 Jiaxi Jiang, Paul Streli, Manuel Meier, Christian Holz (2024). "EgoPoser: Robust Real-Time Egocentric Pose Estimation from Sparse and Intermittent Observations Everywhere". Computer Vision, ECCV 2024, Lecture Notes in Computer Science, pp. 277-294. Springer. https://doi.org/10.1007/978-3-031-72627-9_16. Retrieved 2026-11-01.
  3. ↑ 3.00 3.01 3.02 3.03 3.04 3.05 3.06 3.07 3.08 3.09 3.10 3.11 3.12 Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, Artsiom Sanakoyeu (2023). "Avatars Grow Legs: Generating Smooth Human Motion from Sparse Tracking Inputs with Diffusion Model". IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2023). https://doi.org/10.1109/CVPR52729.2023.00054. Retrieved 2026-11-01.
  4. ↑ 4.00 4.01 4.02 4.03 4.04 4.05 4.06 4.07 4.08 4.09 4.10 Han Feng, Wenchao Ma, Quankai Gao, Xianwei Zheng, Nan Xue, Huijuan Xu (2024). "Stratified Avatar Generation from Sparse Observations". IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024). https://doi.org/10.1109/CVPR52733.2024.00023. Retrieved 2026-11-01.
  5. ↑ 5.00 5.01 5.02 5.03 5.04 5.05 5.06 5.07 5.08 5.09 5.10 5.11 5.12 5.13 5.14 5.15 Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, Christian Holz (2022). "AvatarPoser: Articulated Full-Body Pose Tracking from Sparse Motion Sensing". Computer Vision, ECCV 2022, Lecture Notes in Computer Science, pp. 443-460. Springer. https://doi.org/10.1007/978-3-031-20065-6_26. Retrieved 2026-11-01.
  6. ↑ 6.00 6.01 6.02 6.03 6.04 6.05 6.06 6.07 6.08 6.09 Xiaozheng Zheng, Zhuo Su, Chao Wen, Zhou Xue, Xiaojie Jin (2023). "Realistic Full-Body Tracking from Sparse Observations via Joint-Level Modeling". IEEE/CVF International Conference on Computer Vision (ICCV 2023). https://doi.org/10.1109/ICCV51070.2023.01349. Retrieved 2026-11-01.
  7. ↑ 7.0 7.1 Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, Thomas J. Cashman (2022). "FLAG: Flow-based 3D Avatar Generation from Sparse Observations". IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2022). https://doi.org/10.1109/CVPR52688.2022.01290. Retrieved 2026-11-01.
  8. ↑ 8.0 8.1 8.2 8.3 8.4 8.5 Alexander Winkler, Jungdam Won, Yuting Ye (2022). "QuestSim: Human Motion Tracking from Sparse Sensors with Simulated Avatars". SIGGRAPH Asia 2022 Conference Papers. ACM. https://doi.org/10.1145/3550469.3555411. Retrieved 2026-11-01.
  9. ↑ 9.0 9.1 "Movement SDK for Unity - Overview". Meta Horizon OS Developers. Meta Platforms. 2026-09-30. https://developers.meta.com/horizon/documentation/unity/move-overview/. Retrieved 2026-11-01.
  10. ↑ 10.0 10.1 "Create More Natural Movements Using Inside-Out Body Tracking and Generative Legs". Meta for Developers. Meta Platforms. 2023-12-20. https://developers.meta.com/horizon/blog/inside-out-body-tracking-and-generative-legs/. Retrieved 2026-11-01.
  11. ↑ 11.0 11.1 11.2 Dongseok Yang, Doyeon Kim, Sung-Hee Lee (2021). "LoBSTr: Real-time Lower-body Pose Prediction from Sparse Upper-body Tracking Signals". Computer Graphics Forum, vol. 40, no. 2, pp. 265-275. https://doi.org/10.1111/cgf.142631. Retrieved 2026-11-01.
  12. ↑ Jose Luis Ponton, Haoran Yun, Andreas Aristidou, Carlos Andujar, Nuria Pelechano (2023). "SparsePoser: Real-time Full-body Motion Reconstruction from Sparse Data". ACM Transactions on Graphics, vol. 43, no. 1. https://doi.org/10.1145/3625264. Retrieved 2026-11-01.
  13. ↑ 13.0 13.1 13.2 13.3 13.4 Peng Dai, Yang Zhang, Tao Liu, Zhen Fan, Tianyuan Du, Zhuo Su, Xiaozheng Zheng, Zeming Li (2024). "HMD-Poser: On-Device Real-time Human Motion Tracking from Scalable Sparse Observations". IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024). https://doi.org/10.1109/CVPR52733.2024.00089. Retrieved 2026-11-01.
  14. ↑ Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, Michael J. Black (2015). "SMPL: A Skinned Multi-Person Linear Model". ACM Transactions on Graphics, vol. 34, no. 6. https://doi.org/10.1145/2816795.2818013. Retrieved 2026-11-01.
  15. ↑ 15.0 15.1 15.2 "Body Tracking for Movement SDK for Unity". Meta Horizon OS Developers. Meta Platforms. 2026-10-07. https://developers.meta.com/horizon/documentation/unity/move-body-tracking/. Retrieved 2026-11-01.
  16. ↑ Sadegh Aliakbarian, Fatemeh Saleh, David Collier, Pashmina Cameron, Darren Cosker (2023). "HMD-NeMo: Online 3D Avatar Motion Generation From Sparse Observations". IEEE/CVF International Conference on Computer Vision (ICCV 2023). https://doi.org/10.1109/ICCV51070.2023.00882. Retrieved 2026-11-01.
  17. ↑ Karan Ahuja, Eyal Ofek, Mar Gonzalez-Franco, Christian Holz, Andrew D. Wilson (2021). "CoolMoves: User Motion Accentuation in Virtual Reality". Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 5, no. 2. https://doi.org/10.1145/3463499. Retrieved 2026-11-01.
  18. ↑ Songpengcheng Xia, Yu Zhang, Zhuo Su, Xiaozheng Zheng, Zheng Lv, Guidong Wang, Yongjie Zhang, Qi Wu, Lei Chu, Ling Pei (2025). "EnvPoser: Environment-aware Realistic Human Motion Estimation from Sparse Observations with Uncertainty Modeling". IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2025). https://doi.org/10.1109/CVPR52734.2025.00178. Retrieved 2026-11-01.
  19. ↑ Andrea Dittadi, Sebastian Dziadzio, Darren Cosker, Ben Lundell, Thomas J. Cashman, Jamie Shotton (2021). "Full-Body Motion from a Single Head-Mounted Device: Generating SMPL Poses from Partial Observations". IEEE/CVF International Conference on Computer Vision (ICCV 2021). https://doi.org/10.1109/ICCV48922.2021.01148. Retrieved 2026-11-01.
  20. ↑ Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbeláez, Ali Thabet, Artsiom Sanakoyeu (2023). "BoDiffusion: Diffusing Sparse Observations for Full-Body Human Motion Synthesis". IEEE/CVF International Conference on Computer Vision Workshops (ICCVW 2023). https://doi.org/10.1109/ICCVW60793.2023.00456. Retrieved 2026-11-01.
  21. ↑ Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, Alexander Winkler (2023). "QuestEnvSim: Environment-Aware Simulated Motion Tracking from Sparse Sensors". SIGGRAPH 2023 Conference Proceedings. ACM. https://doi.org/10.1145/3588432.3591504. Retrieved 2026-11-01.
  22. ↑ Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, Michael J. Black (2019). "AMASS: Archive of Motion Capture as Surface Shapes". IEEE/CVF International Conference on Computer Vision (ICCV 2019). https://doi.org/10.1109/ICCV.2019.00554. Retrieved 2026-11-01.
  23. ↑ 23.0 23.1 23.2 Lingni Ma, Yuting Ye, Fangzhou Hong, Vladimir Guzov, Yifeng Jiang, et al. (2024). "Nymeria: A Massive Collection of Multimodal Egocentric Daily Motion in the Wild". Computer Vision, ECCV 2024, Lecture Notes in Computer Science, pp. 445-465. Springer. https://doi.org/10.1007/978-3-031-72691-0_25. Retrieved 2026-11-01.
  24. ↑ "Dataset Copyright License for non-commercial scientific research purposes". AMASS. Max Planck Institute for Intelligent Systems. https://amass.is.tue.mpg.de/license.html. Retrieved 2026-11-01.
  25. ↑ "New Day, New You: Avatars are More Expressive and Customizable Starting Today". Meta Quest Blog. Meta Platforms. 2021-04-23. https://www.meta.com/blog/new-day-new-you-avatars-are-more-expressive-and-customizable-starting-today/. Retrieved 2026-11-01.
  26. ↑ 26.0 26.1 26.2 "Unlock High-Fidelity Movements and Enhanced Interactions on Meta Quest 3". Meta for Developers. Meta Platforms. 2023-09-28. https://developers.meta.com/horizon/blog/movements-inputs-interactions-meta-connect-2023-quest-developers-presence-platform/. Retrieved 2026-11-01.
  27. ↑ 27.0 27.1 27.2 David Heaney (2023-09-28). "Quest 3 Will Get Inside-Out Upper Body Tracking & AI Leg Estimation Later This Year". UploadVR. https://uploadvr.com/quest-3-body-tracking. Retrieved 2026-11-01.
  28. ↑ 28.0 28.1 Tomislav Bezmalinovic (2023-12-22). "Upper body tracking for Quest 3 is here and you can try it out". MIXED. https://mixed-news.com/en/quest-3-upper-body-tracking-dodge-arcade-demo/. Retrieved 2026-11-01.
  29. ↑ David Heaney (2024-03-14). "Virtual Desktop Can Now Use Quest 3's Inside-Out Body Tracking To Emulate Worn Vive Trackers". UploadVR. https://www.uploadvr.com/virtual-desktop-quest-3-inside-out-body-tracking-vive-tracker-emulation/. Retrieved 2026-11-01.
  30. ↑ "Unity-Movement v83.0.0". GitHub. Meta Platforms (oculus-samples). 2025-12-16. https://github.com/oculus-samples/Unity-Movement/releases/tag/v83.0.0. Retrieved 2026-11-01.
  31. ↑ David Heaney (2024-08-20). "Pico Motion Trackers Add Leg Tracking And Use A Clever Calibration Technique". UploadVR. https://uploadvr.com/pico-motion-trackers-leg-tracking. Retrieved 2026-11-01.
  32. ↑ "About mocopi". mocopi documentation. Sony Corporation. https://www.sony.co.jp/en/Products/mocopi-dev/en/documents/Home/Aboutmocopi.html. Retrieved 2026-11-01.
Retrieved from "https://vrarwiki.com/index.php?title=Full-body_pose_estimation_from_sparse_tracking&oldid=53293"