Transformer IMU Calibrator: Dynamic On-body IMU Calibration for Inertial Motion Capture

Chengxu Zuo; Feng Xu; Jiawei Huang; Rui Cao; Shihui Guo; Xiangren Shi; Xiao Jiang; Xinyu Yi; Yipeng Qin; Yuan Yao

REVIEW 4 major objections 6 minor 3 references

Transformer IMU Calibrator: Dynamic On-body IMU Calibration for Inertial Motion Capture

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that the two error sources that corrupt body-worn inertial sensors — coordinate drift and wearing offset — can be estimated continuously from a short window of IMU readings, replacing the one-time static calibration and…

desk verdict A genuine new idea with solid in-distribution validation, but the headline claim overreaches beyond the training envelope. read the letter →

arxiv 2506.10580 v1 pith:EUR3P2E5 submitted 2025-06-12 cs.GR cs.CV

Chengxu Zuo , Jiawei Huang , Xiao Jiang , Yuan Yao , Xiangren Shi , Rui Cao , Xinyu Yi , Feng Xu

show 2 more authors

Shihui Guo Yipeng Qin

This is my paper

classification cs.GRcs.CV

keywords inertialmotioncapturedynamiccalibrationcoordinatedriftmeasurementoffsetTransformersparseIMUrotationdiversitysynthetictrainingdata

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Inertial motion capture normally requires a calibration pose at startup and assumes the calibration stays valid for the whole session, but in practice IMU coordinate drift and wearing offsets change over time. This paper tries to replace that long-term static assumption with a dynamic one: estimate the drift $R_{G'G}$ and the offset $R_{BS}$ continuously, from the last few seconds of motion, without any explicit calibration step. The method learns the mapping from IMU readings to these matrices with a Transformer trained on synthetic data, and uses a rotation-diversity trigger so the network only acts when the motion is informative. If the claim holds, sparse-IMU motion capture becomes more user-friendly — no T-pose ritual — and stays accurate over long sessions, cutting average orientation error from 49.98 degrees to 15.20 degrees on their one-hour real dataset.

What carries the argument

The two calibration matrices are $R_{G'G}$, the coordinate drift rotating the global frame due to magnetic interference and gyro integration error, and $R_{BS}$, the sensor-to-bone measurement offset from imperfect wearing. The argument runs through Eq. (1), which writes the raw IMU reading as $R_{\mathrm{IMU}}(t)=R_{G'G}(t)\,R_{GB}(t)\,R_{BS}(t)$, and through two relaxed assumptions: the Short-term Static Assumption (the matrices barely change inside a window) and the Short-term Diversity Assumption (the IMU orientations in the window are not all the same). These shrink the solution space enough that a Transformer encoder with two pooling-mapping heads — the TIC network — can regress the matrices from $n$ frames of orientation and acceleration. The acceleration input is what lets the network tell the two matrices apart, because acceleration depends on $R_{G'G}$ but not on $R_{BS}$. A trigger based on rotation diversity discretizes Euler angles into a $24 \times 12 \times 24$ grid and only applies the network's output when enough grid cells have been visited.

What would settle it

Take the six-sensor setup and deliberately induce a step change in calibration parameters — for example, re-seat the forearm sensor mid-session or pass a strong magnet near the hip — then check whether TIC's estimated $R_{G'G}$ follows the jump or lags until the next diverse window; a test where the true drift steps by more than the simulated ±60 degrees of yaw and the calibrated pose error spikes would show the short-term static assumption does not hold.

Watch

Extended reading notes

Core claim

The central claim is that the two unknown error matrices — the coordinate drift $R_{G'G}(t)$ and the measurement offset $R_{BS}(t)$ — can be recovered in real time from a short window of raw IMU readings, even though the measurement equation $R_{\mathrm{IMU}}(t)=R_{G'G}(t)\,R_{GB}(t)\,R_{BS}(t)$ is underdetermined. The paper shows this becomes solvable once two relaxed assumptions replace the usual long-term static assumption: the matrices change negligibly inside a short window, and the window's IMU orientations are diverse. A Transformer network (TIC) trained on synthetic paired data learns the mapping from $n$ frames of orientation and acceleration to the two matrices, and a rotation-diversity trigger decides when the mapping's output is reliable enough to apply. On a one-hour real dataset, average orientation measurement error falls from 49.98 degrees with static calibration to 15.20 degrees with TIC, and six existing sparse-IMU pose estimators all improve.

Load-bearing premise

The load-bearing premise is that real drift and offset stay inside the ranges used to synthesize the training data (yaw drift within ±60 degrees for non-root sensors, offsets within ±45 degrees per axis) and change slowly enough inside a short window for the short-term static assumption to hold; if a user's error jumps suddenly or falls outside those ranges, the method has no reliable output.

Editorial extensions

If this is right

Sparse-IMU capture can start from any pose and self-correct after a short burst of diverse motion, removing the T-pose and heading-reset requirements of static calibration.
Long sessions stay accurate: without TIC the six tested estimators degrade to angular errors around 30–37 degrees, while with TIC they stay around 15–19 degrees.
The calibrator is fast enough for continuous background use: rotation-diversity computation takes 2.5 ms and network inference 2 ms, supporting updates every second.
Imperfect initial calibration poses are absorbed automatically, including forearm offsets averaging about 29–33 degrees in the collected data.
The same pipeline transfers to consumer-grade IMUs in phones, watches, and earbuds, reducing average orientation measurement error from 23.61 degrees to 18.95 degrees on the IMUPoser dataset.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The authors do not test the failure boundary directly, but the training ranges in Table 8 imply that a user whose drift exceeds the simulated yaw range of ±60 degrees (or offsets beyond ±45 degrees) will receive estimates the model was never trained to produce; a follow-up could map OME as a function of the true drift magnitude to find the practical envelope.
Because the network outputs incremental updates $\Delta R_{G'G}$ and $\Delta R_{BS}$ rather than absolute matrices, the same calibrator could be stacked after any existing static calibration routine and used to correct only the residual error, not just replace the routine entirely.
A low-activity user (office work, watching TV) will starve the diversity trigger, so a practical deployment would likely pair this method with a vision- or magnetometer-based fallback during sedentary periods — an extension the paper does not explore.
The synthetic-to-real recipe suggests a cheap porting path: re-sample the Euler-angle ranges to match a new sensor population (e.g., noisier consumer IMUs or looser attachment) and fine-tune, which could extend the method beyond the six-body configuration tested here.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's my take: the paper delivers a genuine new idea—replacing the once-per-use static calibration of sparse IMU mocap with a learned continuous estimator of coordinate drift and sensor offset—and the core validation is solid enough that it deserves referee time. I don't think the central mechanism is broken; the soft spots are about how far the claim extends.

What it does well: breaking the long-term static assumption is a real departure from Seel/Taetz/Mueller, which estimate fixed session parameters. The synthetic data synthesis is careful, including the gravity-leakage model in acceleration, and the ablation shows the acceleration channel (ACCA) is load-bearing for separating R_GG' from R_BS. The real test set—five subjects, one hour each, optical ground truth—is a useful resource, and the reproducibility check across ten training runs (Table 13) is a nice touch. The limitations section is unusually honest: it admits failure under large sudden changes, low activity, and irregular motion, and it flags the global-yaw drift problem.

Now the soft spots. The biggest one, I agree with the stress-test note: the synthetic training envelope is effectively tuned to the test distribution. The ±45° for R_BS is justified by saying the observed T-pose errors are 'well covered'—which means the evaluation is inside the training box, and there is no evidence the transformer extrapolates to larger offsets or to the mixed-axis drift shown in Fig. 13. Likewise, each synthetic window uses a constant transform, while real drift at 0.309°/s changes R_GG' by about 2.6° inside one 8.53-s window; the 'negligibly' in Assumption 2 is strained at the fast end of the authors' own data. This doesn't sink the paper, but it does mean the 'implicit calibration for arbitrary use' headline is overbroad. The paper is honest in its limitations, but the abstract's 'first to enable long-term and accurate motion capture using sparse IMUs' needs to be qualified to the ego-yaw frame and the tested envelope.

A couple of smaller things: Table 6 confirms translation in the SMPL world frame is worse with TIC, so the improvement is in the ego-yaw frame only. The RD trigger ablation is thin—0.28° OME difference—though the qualitative plot shows it prevents oscillation in low-diversity windows. Only five subjects, and Table 13 gives training variance but not subject-level variance; I'd want that. Also, hyperparameters like the update interval t were selected on the test set (Table 11).

Who is this for? People working on sparse inertial mocap, VR, or fitness tracking. A serious editor should send it out; the reviewer's job is to push for scope clarification, subject-level variance, and an OOD test (e.g., simulate larger drift or within-window drift and see if the model degrades gracefully). I'd accept it after those revisions, and I'd cite it for the dataset and the ego-yaw decomposition.

Referee Report

4 major / 6 minor

Summary. This paper proposes Transformer IMU Calibrator (TIC), a dynamic calibration method for sparse inertial motion capture. The method replaces the standard assumption that the coordinate drift R_G'G and measurement offset R_BS are constant for an entire session with two weaker assumptions: that these quantities change negligibly within a short time window and that the IMU readings within that window are diverse. A Transformer is trained on synthetic AMASS/DIP data with randomly sampled R_G'G and R_BS to regress residual updates from a 256-frame window of six IMU orientations and accelerations, and a rotation-diversity trigger gates which updates are applied. The authors evaluate on a new one-hour, five-subject dataset with optical ground truth, reporting a reduction in average orientation measurement error from 49.98 degrees to 15.20 degrees, and they show improvements in six downstream sparse-IMU pose estimators, along with ablations and a 10-run reproducibility study.

Significance. If the method generalizes beyond its training envelope, it is a valuable step: it removes the explicit T-pose calibration step, addresses long-term drift, and contributes a public dataset and code. The central result is not simply re-fitting the answer, because the ground-truth R_G'G and R_BS are derived from independently measured optical motion capture (Eq. 7), and the ablation study gives clear evidence that the acceleration auxiliary input matters (Table 4). The 10-run reproducibility study (Appendix J) and evaluation on consumer-grade IMUs (Table 7) and Xsens IMUs (Table 15) are additional strengths. The main weakness is that generalization is currently conditional on the synthetic training distribution and on the validity of Assumption 2, and several claims in the abstract and conclusion are stronger than the evidence supports.

major comments (4)

[§5.1, Appendix A (Table 8), Appendix C (Table 9)] The synthetic training distribution is effectively tuned to the evaluation distribution. Appendix C states that the observed forearm T-pose errors (32.86 +/- 11.43 degrees) are 'well covered' by the +/-45 degree R_BS range used in Appendix A, so the simulation box is not an independent prior over deployment conditions. The paper provides no out-of-distribution evaluation, e.g., with R_BS beyond +/-45 degrees, with mixed-axis R_G'G drift as seen for IMU 2 in Fig. 13, or with residual deltas outside the prescribed ranges. Because the headline claim is implicit calibration for arbitrary use, the behavior outside this box is load-bearing. Please add synthetic out-of-distribution tests or otherwise quantify the coverage of realistic deployment conditions.
[§4.3 (Algorithm 2), §5.1, Fig. 13] Assumption 2 is never directly tested. Training draws one constant R_G'G and R_BS per 256-frame sequence (Section 5.1), while deployment estimates residual deltas on windows in which the underlying R_G'G is drifting. At the observed drift rate of about 0.309 degrees/s (Fig. 13), R_G'G changes by roughly 2.6 degrees within one 8.53-s window, which is not negligible relative to the 9.18-degree mean R_G'G error reported in Table 4. The paper should include a controlled synthetic experiment with within-window parameter drift, or show that the residual process seen in deployment lies within the constant-per-window training distribution.
[§5.7, Table 6] The claim that dynamic calibration significantly enhances translation estimation is supported only in the ego-yaw frame. In the SMPL (fixed global) frame, TIC increases translation error at every reported window (1s: 18.03 vs 14.09 cm; 10s: 49.50 vs 41.69 cm). Since the abstract and conclusion promise 'long-term and accurate motion capture using sparse IMUs,' the paper should either restrict the claim to body-relative pose or explain why the degradation in the fixed global frame is acceptable for the target application. This caveat should appear in the abstract and conclusion, not only in a sentence inside Section 5.7.
[§5.4, Table 2; §6] The headline OME/AME improvement is demonstrated on a single one-hour dataset whose measured parameters fall inside the synthetic box, while Section 6 concedes failure under large sudden changes, low-activity windows, and irregular motions. The paper should provide a per-subject, per-activity, or per-window breakdown of the errors so that readers can assess how often the method operates in regimes where it is expected to help. Without such a breakdown, the average 49.98-to-15.20 reduction may mask substantial portions of time where TIC is not triggered or is less accurate.

minor comments (6)

[§2.1] There are typographical errors in the related-work section: 'arbirtary' should be 'arbitrary' and 'hight efficiency' should be 'high efficiency.'
[Table 1 and §5.4] The dataset naming is inconsistent: Table 1 uses 'DS AMS,' 'DS DIP,' and 'DS TIC,' while the text uses 'DSTIC' and 'DS DIP.' Please standardize the names.
[§4.3, Algorithm 2] The buffer semantics of Algorithm 2 are ambiguous. If B_n is a fixed-length queue, the condition |B_n|==n and the clear operation are at odds; if B_n is a growing list capped at n, the condition may never be met because the 1-s timing signal does not align with the 256-frame fill time. Please clarify whether the buffer is a sliding window and how the timing signal interacts with it.
[Appendix G] There is a typo in the training details: 'The a_IMU were divide by 30' should read 'The a_IMU were divided by 30,' and there is a stray comma after the citation '[Zhou et al. 2019].'
[Fig. 13 and Table 2] The text refers to 'the IMU on the right hand (IMU 2),' while Table 2 and the data-collection description consistently call this location the right forearm. Please use consistent body-location naming.
[§5.8, Table 7] Table 7 shows that for the right front pocket sensor, OME with TIC (10.64 degrees) is worse than without TIC (9.18 degrees). The claim that TIC 'significantly reduced OME in the IMUPoser dataset' should be accompanied by a per-sensor caveat.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the TIC mapping is trained on synthetic paired data and tested against independent optical ground truth; the test-set-tuned synthetic envelope and hyperparameters are generalization risks, not circular reductions.

full rationale

The claimed derivation chain is not circular. The network is trained with Eq. 6 to regress R_G'G(n) and R_BS(n) from synthetic IMU sequences generated by applying random Euler-angle rotations to AMASS/DIP data (Sec. 5.1, Appendix A, Table 8), and it is evaluated on DS_TIC, whose ground-truth drift and offset are computed independently from optical markers via Eq. 7 (Sec. 5.2). The test labels therefore do not enter the training objective, and the w-TIC vs w/o-TIC comparisons (Tables 2 and 3) are externally grounded improvements, not constructions equivalent to the inputs. The main caveats are three test-set-informed design choices: Appendix C states that the observed T-pose errors are 'well covered' by the R_BS sampling range, meaning the simulator's envelope (Table 8) is deliberately matched to the evaluation distribution; Table 11's update interval t=1s is chosen from errors measured on DS_TIC; and Table 16's delta-parameterization is chosen on the same set. These choices weaken generalization claims and make the evaluation partly in-distribution, but they do not make the predicted OME/AME values equal to fitted parameters by construction; the network still must learn a nontrivial mapping and fails on out-of-envelope conditions, as the Limitations section (Sec. 6) concedes for sudden changes, low activity, and irregular motions. No load-bearing self-citation or imported uniqueness theorem is used. Overall, this is a correct non-circular finding with minor generalization-risk caveats, hence score 1.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The method rests on two modeling assumptions (short-term constancy and diversity), a synthetic training distribution that encodes the authors' prior about real drift and offsets, and the ego-yaw frame that deliberately excludes global yaw drift. The main free parameters are the synthetic sampling ranges and the trigger thresholds, all chosen by hand or by validation sweeps. No new physical entities are proposed; the ego-yaw frame is a coordinate convention with clearly stated scope.

free parameters (6)

R_BS synthetic sampling range = U(-45,45) degrees on x, y, z axes (Table 8)
Training distribution of measurement offset; the model cannot estimate offsets well outside this range.
R_GG' synthetic sampling ranges = root: U(-20,20), 0, U(-20,20); non-root: U(-20,20), U(-60,60), U(-20,20) (Table 8)
Training distribution of coordinate drift; the yaw range is informed by drift analysis but hand-chosen.
Per-joint rotation diversity threshold T_R = 15 (hip), 25 (head), 30 (left/right lower leg and left forearm), 50 (right forearm) (Table 10)
Trigger thresholds tuned on the KIT subset of AMASS to balance accuracy and trigger frequency.
Update interval t = 1 s (Table 11)
Chosen from a sweep on the authors' dataset; below 2 s no significant difference, performance degrades at 20 s.
Window length n = 256 frames (8.53 s at 30 Hz)
Fixed history length for input; not swept across values.
Euler angle discretization step = 15 degrees (24x12x24 grid)
Chosen for computational feasibility of the rotation diversity computation.

assumptions (6)

domain assumption Assumption 2: R_GG' and R_BS change negligibly within a short time window.
The entire estimation framework is built on this; large sudden changes break it, as the Limitations section concedes.
domain assumption Assumption 3: IMU readings within the window are diverse.
The RD trigger enforces it, but the learned mapping also requires diversity to disambiguate; low-activity scenarios fail, as the Limitations section concedes.
domain assumption The world drift R_W'G decomposes into a non-solvable yaw component R_W'G' and a solvable component R_GG', defining the ego-yaw frame.
Appendix D: this decomposition sets the scope of calibration; global yaw drift is excluded by construction, matching the worse SMPL-frame translation in Table 6.
domain assumption Uniform random Euler-angle sampling of R_GG' and R_BS over the ranges in Table 8 covers real-world drift and offset distributions.
Section 5.1 and Appendix A: the model inherits this prior; real mixed-axis drift and T-pose errors near the range boundaries are only partially validated by the drift analysis in Appendix B.
domain assumption The hardware-level GA removal model with g_bias fixed to the true gravity vector (Eq. 10).
Used to synthesize acceleration with GA leakage; assumes the IMU's gravity removal is perfect except for the drift.
domain assumption The human motion prior in AMASS and DIP is representative of real user motion.
The network implicitly relies on pose priors to disambiguate calibration parameters; irregular motions in the demo video caused errors, as the Limitations section concedes.

invented entities (1)

ego-yaw coordinate system
purpose: Defines the global frame with zero roll/pitch and body-synchronized yaw, making the non-solvable global yaw drift invisible to the calibration task.
A definitional modeling construct, not an independently testable entity; it bounds the method to pose-relative accuracy and excludes global yaw correction, matching the worsened SMPL-frame translation in Table 6.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Transformer IMU Calibrator: Dynamic On-body IMU Calibration for Inertial Motion Capture." pith.science (2026). https://pith.science/paper/EUR3P2E5

@misc{pith2026250610580,
  author       = {Pith},
  title        = {Pith review of: Transformer IMU Calibrator: Dynamic On-body IMU Calibration for Inertial Motion Capture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUR3P2E5}},
  note         = {Machine review of arXiv:2506.10580}
}

read the original abstract

In this paper, we propose a novel dynamic calibration method for sparse inertial motion capture systems, which is the first to break the restrictive absolute static assumption in IMU calibration, i.e., the coordinate drift RG'G and measurement offset RBS remain constant during the entire motion, thereby significantly expanding their application scenarios. Specifically, we achieve real-time estimation of RG'G and RBS under two relaxed assumptions: i) the matrices change negligibly in a short time window; ii) the human movements/IMU readings are diverse in such a time window. Intuitively, the first assumption reduces the number of candidate matrices, and the second assumption provides diverse constraints, which greatly reduces the solution space and allows for accurate estimation of RG'G and RBS from a short history of IMU readings in real time. To achieve this, we created synthetic datasets of paired RG'G, RBS matrices and IMU readings, and learned their mappings using a Transformer-based model. We also designed a calibration trigger based on the diversity of IMU readings to ensure that assumption ii) is met before applying our method. To our knowledge, we are the first to achieve implicit IMU calibration (i.e., seamlessly putting IMUs into use without the need for an explicit calibration process), as well as the first to enable long-term and accurate motion capture using sparse IMUs. The code and dataset are available at https://github.com/ZuoCX1996/TIC.

Figures

Figures reproduced from arXiv: 2506.10580 by the authors.

**Figure 1.** Live demonstration of our dynamic calibration method against the conventional static calibration method. Our method provides a user-friendly experience (w/o IMU heading reset and T-Pose) and ensures long-term robustness for inertial motion capture. ACM Trans. Graph., Vol. 44, No. 4, Article . Publication date: August 2025. arXiv:2506.10580v1 [cs.GR] 12 Jun 2025 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 3.** The rationale of our dynamic calibration. The [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Our dynamic calibration workflow. With real-time IMU inputs, the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (13 more)

**Figure 5.** Figure 5: Our data collection system. The absolute IMU orientation and ac [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 7.** Figure 7: Qualitative evaluation on of Rotation Diversity ( [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: Qualitative evaluation on 𝑅𝐵𝑆 error. The visualized 𝑅𝐵𝑆 error and positional error are the average values of left and right forearm. The examples are from subject 5 in DSTIC [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: Visualization of global translation tracking on DSTIC in the SMPL frame and ego-yaw frame, unit: m. coordinate system is markedly lower than when it is not used. This improvement is attributed to the fact that translation estimation relies on reliable pose estimation r…

**Figure 10.** Figure 10: Evaluation on complex and irregular motions in the demo video. We observed that such scenarios could lead to inaccurate calibration, as these types of motions are underrepresented in existing motion datasets. Fortunately, our approach can automatically recover and cor…

**Figure 11.** Figure 11: Illustration of gravitational acceleration (GA) leakage in hardware [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]

**Figure 14.** Figure 14: T-Pose error in real-world usage [PITH_FULL_IMAGE:figures/full_fig_p011_14.png]

**Figure 13.** Figure 13: Visualization of IMU drifting in real-world dataset (s1). We use [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]

**Figure 15.** Figure 15: Decompose of the coordinate drift based on the ego-yaw coordinate [PITH_FULL_IMAGE:figures/full_fig_p011_15.png]

**Figure 17.** Figure 17: Four examples of daily activities from KIT dataset. Motion labels [PITH_FULL_IMAGE:figures/full_fig_p012_17.png]

**Figure 16.** Figure 16: The impact of rotation diversity on OME. [PITH_FULL_IMAGE:figures/full_fig_p012_16.png]

**Figure 18.** Figure 18: Architecture of TIC network. The network uses three stacked Trans [PITH_FULL_IMAGE:figures/full_fig_p013_18.png]

**Figure 19.** Figure 19: An example of our dynamic calibration on Xsens IMUs. We visu [PITH_FULL_IMAGE:figures/full_fig_p014_19.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages

[2009]

Julien Favre, BM Jolles, Rachid Aissaoui, and K Aminian

Functional calibration procedure for 3D knee joint angle description using inertial sensors.Journal of biomechanics42, 14 (2009), 2330–2335. Julien Favre, BM Jolles, Rachid Aissaoui, and K Aminian. 2008. Ambulatory measure- ment of 3D knee joint angle.Journal of biomechanics41, 5 (2008), 1029–1035. Aparna Harindranath and Manish Arora. 2024. A systematic ...

arXiv 2009
[2021]

ACM Interact

CoolMoves: User Motion Accentuation in Virtual Reality.Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.5, 2 (2021). Rayan Armani, Changlin Qian, Jiaxi Jiang, and Christian Holz. 2024. Ultra Inertial Poser: Scalable Motion Capture and Tracking from Sparse Inertial Sensors and Ultra-Wideband Ranging. InSIGGRAPH 2024 Conference Papers. Association for C...

work page 2021
[2024]

Vasco Xu, Chenfeng Gao, Henry Hoffmann, and Karan Ahuja

Fast Human Motion reconstruction from sparse inertial measurement units considering the human shape.Nature Communications15, 1 (2024), 2423. Vasco Xu, Chenfeng Gao, Henry Hoffmann, and Karan Ahuja. 2024. MobilePoser: Real-Time Full-Body Pose Estimation and 3D Human Translation from IMUs in Mobile Consumer Devices. InProceedings of the 37th Annual ACM Symp...

work page 2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [2009]

Julien Favre, BM Jolles, Rachid Aissaoui, and K Aminian

Functional calibration procedure for 3D knee joint angle description using inertial sensors.Journal of biomechanics42, 14 (2009), 2330–2335. Julien Favre, BM Jolles, Rachid Aissaoui, and K Aminian. 2008. Ambulatory measure- ment of 3D knee joint angle.Journal of biomechanics41, 5 (2008), 1029–1035. Aparna Harindranath and Manish Arora. 2024. A systematic ...

arXiv 2009

[2] [2021]

ACM Interact

CoolMoves: User Motion Accentuation in Virtual Reality.Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.5, 2 (2021). Rayan Armani, Changlin Qian, Jiaxi Jiang, and Christian Holz. 2024. Ultra Inertial Poser: Scalable Motion Capture and Tracking from Sparse Inertial Sensors and Ultra-Wideband Ranging. InSIGGRAPH 2024 Conference Papers. Association for C...

work page 2021

[3] [2024]

Vasco Xu, Chenfeng Gao, Henry Hoffmann, and Karan Ahuja

Fast Human Motion reconstruction from sparse inertial measurement units considering the human shape.Nature Communications15, 1 (2024), 2423. Vasco Xu, Chenfeng Gao, Henry Hoffmann, and Karan Ahuja. 2024. MobilePoser: Real-Time Full-Body Pose Estimation and 3D Human Translation from IMUs in Mobile Consumer Devices. InProceedings of the 37th Annual ACM Symp...

work page 2024