Pith. sign in

REVIEW 3 major objections 6 minor 57 references

Sen-Cap: Sensor-Flexible and Noise-Resilient Human Motion Capture via LiDAR-Camera Integration

T0 review · 3 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Sen-Cap claims that one unified model can reconstruct full SMPL pose, shape, and global trajectory from an arbitrary mix of uncalibrated LiDARs and cameras, without retraining.

desk verdict A serious systems paper with real benchmark gains and a genuinely useful noise-robust tracker, but the headline 'calibration-free arbitrary deployment' claim is not yet demonstrated because alignment is rotation-only and unseen LiDAR extrinsics are never tested. read the letter →

arxiv 2608.02285 v1 pith:67SKMAV6 submitted 2026-08-03 cs.CV

classification cs.CV
keywords humanmotioncaptureLiDAR-camerafusioncalibration-freesensorflexibilitySMPLnoiserobustnesstrajectoryestimationcross-modal
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

Sen-Cap's aim is to make multi-sensor 3D human motion capture practical outdoors: it tries to show that explicit calibration between LiDARs and cameras is unnecessary, and that global trajectories can survive severe point-cloud clutter. The paper proposes a single trained model that, at inference, accepts any number and combination of LiDARs and cameras, aligns all LiDAR observations into a human-centric coordinate frame by predicting a rotation per sensor, and fuses them with camera keypoints through a bottleneck attention mechanism. A separate iterative tracker refines global translation by repeatedly moving the point-cloud center toward the estimated pelvis, which stabilizes trajectories under noise. The paper reports that this design outperforms existing camera-only, LiDAR-only, and hybrid methods on the Human-M3 and FreeMotion benchmarks, including under novel camera viewpoints, and that it transfers to other datasets without fine-tuning while running in real time.

What carries the argument

Three mechanisms carry the argument. (1) Human-Centric Space Alignment: a GRU per LiDAR predicts a rotation vector tau_i, and the LiDAR's joint features are transformed via x^H = R(tau_i)^{-1} x^L into one canonical frame, trained with an alignment loss against ground-truth sensor-to-human rotations. (2) Bottleneck Adaptive Fusion: all modality-specific features are concatenated and then aggregated by cross-attention from a single learnable token, which forces sensor contributions to be re-weighted by reliability rather than mixed unrestrictedly. (3) Noise-Resistant Trajectory Tracker (NTT): an offset predictor is applied iteratively, shifting the point cloud toward the predicted root joint

What would settle it

Take a trained model and at inference place a LiDAR at a height, distance, or azimuth far outside the training layout distribution while keeping the person fixed; if per-joint error rises to the level of a system that requires calibration, the learned alignment has not generalized. A cleaner protocol: hold out whole families of sensor layouts during training and test on them, comparing against a calibration-based upper bound.

Watch

Extended reading notes

Core claim

On its own terms, Sen-Cap's central discovery is that the geometric alignment between heterogeneous sensors can be learned as a rotation into a human-centric space rather than estimated as a calibration matrix. Each LiDAR's 3D joint features are rotated by a GRU-predicted rotation R(tau_i)^{-1} into a canonical frame centered on the person; 2D keypoints are encoded into the same space through bounding-box normalization; and all sensor features are squeezed through a single learnable latent token that re-weights unreliable inputs. The global trajectory is then recovered by an iterative refinement that doesn't assume the point-cloud centroid coincides with the pelvis. Under the paper's evaluat

Load-bearing premise

The claim collapses if the network's learned rotation into the human-centric frame does not generalize to sensor poses it has never seen, because no explicit calibration corrects the alignment at test time.

Editorial extensions

If this is right

  • A single trained model can be deployed with one, two, or more LiDARs and cameras in any combination; the same weights handle fewer or additional sensors at test time, and adding sensors generally reduces error.
  • The system degrades gracefully when a sensor is lost or its view is occluded, because training with Sensor Dropout teaches the fusion to lean on whichever sensors remain reliable.
  • Trajectory estimates remain consistent under synthetic point-cloud clutter where standard centroid-normalization drifts, with the largest gains at the highest noise level.
  • Because local pose and global trajectory are computed in a human-centric frame, the method transfers across environments and datasets without fine-tuning, at least within the tested domain gaps.
  • The whole pipeline runs in real time, so the approach is suitable for live capture rather than offline processing.

Reading between the lines

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

  • Because the learned alignment predicts only a rotation, not a translation, the 'human-centric space' is really a shared orientation frame; a sensor placed at a very different distance from the person may still need rough localization, so the flexibility claim is strongest for layouts that resemble the training distribution.
  • The bottleneck fusion re-weights sensors implicitly, but it never estimates per-sensor uncertainty explicitly; a natural extension is to feed the fusion an explicit confidence signal per sensor, which could harden behavior beyond random dropout.
  • The novel-viewpoint protocol changes camera extrinsics while 2D keypoints are bounding-box normalized, so the camera side of the calibration-free claim is partially stress-tested, but LiDAR extrinsics are not varied at test time; a direct experiment would rotate or displace the LiDARs themselves at inference and measure error.
  • The noise evaluation uses synthetic clutter placed around the person; a harder, more realistic test is occluders that move like people, which would pressure both the fusion and the trajectory tracker in a way the current object-noise levels do not.
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.

Referee Report

3 major / 6 minor

Summary. Sen-Cap proposes a single motion-capture model that reconstructs local SMPL pose/shape and global trajectories from a variable, uncalibrated set of LiDARs and cameras. The method has two main components: a Unified Across-Sensor Motion Estimator (UAME), which aligns per-LiDAR 3D joint features into a human-centric space via a predicted rotation and fuses them with camera keypoint features through a bottleneck cross-attention token, and a Noise-Resistant Trajectory Tracker (NTT), which refines global translation by iterative offset prediction. The paper evaluates on FreeMotion, Human-M3, LiDARHuman26M, and RELI11D, reporting state-of-the-art or competitive results, and includes ablations for feature space, fusion strategy, NTT iterations, sensor-count scaling, and synthetic noise levels.

Significance. If validated, the paper addresses a practically important limitation of current hybrid LiDAR-camera mocap: removing explicit calibration at inference and permitting flexible, changing sensor configurations. The empirical evaluation is broad and internally consistent, with four public benchmarks, targeted sensor-flexibility and noise-resilience protocols, and a real-world deployment demo. The central scientific claim, however, rests on the human-centric alignment generalizing to unseen LiDAR extrinsics and on the alignment being a complete rigid transform; the current evidence for this is indirect. Because both issues are addressable with additional analysis and experiments, the contribution is potentially significant but not yet fully established.

major comments (3)
  1. [§3.1, Eqs. (1)-(2)] The alignment equation uses only a rotation: x^Hi_3d = R(τ_i)^{-1} x^Li_3d. A LiDAR-to-human rigid transform has six degrees of freedom. Unless x^Li_3d is explicitly pre-centered or uses a translation-invariant representation, applying only R(τ_i)^{-1} leaves each LiDAR's translational offset unresolved, so LiDARs at different positions cannot be brought into the same human-centric frame. The paper does not state that x^Li_3d is pre-centered, and §3.2 actually treats centroid normalization as problematic for NTT. Please specify the full SE(3) alignment or explicitly state and evaluate any implicit centering/normalization step.
  2. [§4.2, §4.6 and claims in §1/§3] No experiment varies LiDAR extrinsics at test time. The novel-view protocol changes only camera viewpoints, and the 2D keypoints are bbox-normalized, which largely absorbs camera translation/scale. The sensor-flexibility experiments change the number of sensors but keep the deployed sensor layouts fixed. Since the GRU that predicts τ_i is trained with ground-truth sensor-to-human rotations, the ability to infer τ_i for unseen LiDAR poses is unverified. Please add a held-out LiDAR-extrinsics protocol (e.g., train on one set of LiDAR placements, test on another) or substantially weaken the 'arbitrary sensor deployment' claim.
  3. [Table 2] The cross-domain evaluation does not state the training/evaluation protocol for the baselines. The caption says only that 'our approach is trained on AMASS and FreeMotion.' If WHAM, GVHMR, LiveHPS++, FreeCap, etc. are evaluated with different training data or with fine-tuning, the comparison is not apples-to-apples; if they are evaluated zero-shot, that should be stated explicitly. Without this, the 'strong cross-domain performance' claim is difficult to assess.
minor comments (6)
  1. [General] No error bars or repeated-run statistics are reported anywhere. Several differences in Tables 1-4 and Fig. 5 are small (e.g., Ang Err differences of about 1 degree). A single run is insufficient to establish that these differences are not noise; adding standard errors or significance tests would strengthen the empirical claims.
  2. [§3.1, Eq. (2)] The notation mixes τ_i (a rotation vector used in Eq. (1)) with τ^6d_i (a 6D rotation representation). Clarify the relationship between these quantities and how the 6D representation is supervised.
  3. [§4.5, Table 3] The NTT iteration count N is a free parameter, and the main text should state explicitly how N=3 was chosen beyond 'favorable balance.' Also, the table shows that repeat×4 and repeat×5 give further improvements; please report the computational cost of each setting.
  4. [Fig. 5] The legend 'Specialized (XC+YL)' uses X and Y without definition in the caption. Define what 'specialized' means and how the specialized models were trained.
  5. [§3.1] The 2D input alignment is described only as 'encoding the 2D keypoints into a human-centric feature space via the 2D Encoder.' It is unclear whether there is an explicit loss enforcing human-centric consistency for the camera branch, or whether the 2D branch is simply trained end-to-end through the downstream pose loss.
  6. [Abstract and §4.1] The abstract claims real-time operation, but no inference time or runtime comparison is reported. Please provide a runtime number or remove the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: UAME and NTT are supervised modules trained on ground-truth SMPL labels and evaluated on external benchmarks; the calibration-free generalization gap is an empirical-support issue, not a circular reduction.

full rationale

The paper's derivation chain is supervised learning from ground-truth labels, not a self-referential fit. Eq. (1) defines human-centric alignment via a predicted rotation R(τ_i)^{-1}, and Eq. (2) trains that GRU against ground-truth sensor-to-human rotations and joint coordinates; the later pose loss (Eq. 5) and trajectory loss (Eq. 7) are likewise supervised by ground-truth SMPL pose/shape and global translation. None of these predictions reduces by construction to a fitted constant: the GRU must generalize from point-cloud/2D-keypoint inputs to rotations, and NTT's iterative offsets are trained against residuals from GT, with evaluation on held-out benchmarks (Human-M3, FreeMotion, LiDARHuman26M, RELI11D) that provide external falsifiability. The overlapping-author citations (FreeCap, LiveHPS++) supply a matching subroutine and preprocessing choices, but the central claims—human-centric alignment, bottleneck fusion, and iterative trajectory refinement—are independently ablated and benchmarked. The noticeable weakness is evidential rather than circular: the rotation-only alignment of Eq. (1) is under-specified with respect to translation, and no experiment directly varies unseen LiDAR extrinsics, so the 'arbitrary deployment' claim is not fully demonstrated. That is a generalization/correctness concern, not a case where the output is equivalent to the input by definition. Score 0: no significant circularity.

Assumptions & free parameters 7 free parameters · 7 assumptions · 2 invented entities

The central claims rest on supervised learning from calibrated multi-sensor datasets, an implicit learned alignment step, and several hyperparameters chosen with knowledge of the evaluation results. The most consequential free choices are the NTT iteration count (N=3, though N=5 is better in the same table), the unstated sensor-dropout rate that powers the flexibility claim, and the in-house noise levels. No parameter-free derivation is claimed anywhere; the system is an end-to-end trained architecture.

free parameters (7)
  • NTT iteration count N = 3 (main results; 4-5 slightly better in Table 3)
    Table 3 shows monotone improvement with iterations (PST 87.88 to 74.63 mm from 1 to 5); N=3 is selected as 'favorable balance', i.e., tuned with knowledge of the evaluation results. Central to the trajectory-accuracy claim.
  • Trajectory penalty weight lambda_1
    Eq. 7; value not reported in the visible text; regularizes NTT step sizes.
  • Step-size penalty threshold delta
    Eq. 8; chosen by hand; defines the allowed per-iteration offset before penalization.
  • Shape-term weight alpha in L_pose
    Eq. 5; value not reported in the visible text.
  • Sensor dropout masking rate
    Sec. 3.1 'randomly mask a certain percentage of feature tokens'; rate unspecified; this augmentation is the mechanism behind the sensor-flexibility claim (Fig. 5).
  • LiDAR points per frame (FPS) = 256
    x^Li_pc in R^{256x3} (Problem Statement); inherited from LiveHPS++ as the input sampling hyperparameter.
  • Synthetic noise levels L1/L2/L3 = object sizes 0.1-0.5 m / 0.5-1 m
    Sec. 4.6 in-house protocol for the noise-resilience evaluation; levels authored for this paper, not a standard benchmark.
assumptions (7)
  • domain assumption Calibrated training labels with per-sensor ground-truth rotations
    Sec. 3.1 Eq. 2 supervises tau_6d_i with ground-truth tau_6d_i; FreeMotion/Human-M3 supply these. The inference-time 'calibration-free' property is a learned imitation of calibration.
  • domain assumption GRU rotation predictor generalizes to unseen sensor extrinsics
    Eq. 1; the arbitrary-deployment claim depends on this generalization. Only indirect evidence in the visible text (novel camera views change camera, not LiDAR, extrinsics; cross-domain results degrade substantially, Table 2).
  • domain assumption Rotation-only alignment suffices (translation handled implicitly)
    Eq. 1 applies only R(tau_i)^-1; no explicit root-translation alignment step is described, so the construction of the human-centric space is incomplete in the visible text.
  • domain assumption Upstream 2D keypoint detector with bounded error
    Inputs x^Cj_2d are 'normalized 2D keypoints' (Problem Statement); the detector is never named, so camera-pathway noise properties are unspecified.
  • domain assumption SMPL parameterization adequacy
    All pose/shape supervision and evaluation use SMPL; the body-model fit is not questioned.
  • domain assumption Iterative refinement converges
    Eqs. 6-8 assume repeated offset prediction monotonically reduces trajectory error; supported empirically (Table 3) for the tested regime, not generally.
  • domain assumption FreeCap-style matching + memory bank handles multi-person association
    Sec. 3 'we build upon the optimization-based per-frame matching mechanism of FreeCap [50]... introduce a matching memory bank'; details deferred to supplementary; multi-person robustness is assumed.
invented entities (2)
  • Learnable bottleneck fusion token F_token independent evidence
    purpose: Shared query token that compresses all modality features into one latent vector (Eq. 3), enabling variable-sensor-count fusion with reliability weighting.
    Controlled ablation (Table 3) shows bottleneck attention (52.25 mm) beats linear (60.63) and fixed (61.37) fusion; the handle is in-paper rather than an independent benchmark.
  • Matching memory bank
    purpose: Enables real-time online multi-person cross-sensor matching based on FreeCap's optimization matcher.
    Introduced in Sec. 3 with details deferred to supplementary; no visible evaluation isolates its contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sen-Cap: Sensor-Flexible and Noise-Resilient Human Motion Capture via LiDAR-Camera Integration." pith.science (2026). https://pith.science/paper/67SKMAV6

@misc{pith2026260802285,
  author       = {Pith},
  title        = {Pith review of: Sen-Cap: Sensor-Flexible and Noise-Resilient Human Motion Capture via LiDAR-Camera Integration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/67SKMAV6}},
  note         = {Machine review of arXiv:2608.02285}
}
read the original abstract

We propose Sen-Cap, a Sensor-Flexible and Noise-Resilient 3D human motion Capture framework that integrates multi-modal data from LiDAR and camera. While multi-modal sensors provide richer information than single-modal sensors, existing approaches still suffer from two core challenges. First, multi-modal alignment/matching across arbitrarily deployed sensors is typically handled by explicit calibration, which propagates errors under changing viewpoints and in turn constrains deployment to fixed, highly overlapped layouts. Second, prior methods degrade under severe noise or partial sensor failures, which are common in real-world environments. To address these challenges, Sen-Cap introduces a Unified Across-Sensor Motion Estimator that reconstructs local pose and shape in a human-centric space without calibrations between sensors, supporting a flexible number of sensors, as well as a Noise-Resistant Trajectory Tracker that maintains robustness under severe point cloud noise through iterative refinement. These sensor-flexible and noise-resilient features make Sen-Cap more practical in real-world deployment. Notably, operating in real time, Sen-Cap achieves state-of-the-art performance on major metrics on Human-M3 and FreeMotion, as well as strong cross-domain performance on LiDARHuman26M and RELI11D. This combination of flexibility and robustness opens new opportunities for motion capture in real-world scenarios, e.g. sports analytics, field robotics, and large-scale immersive environments.

Figures

Figures reproduced from arXiv: 2608.02285 by the authors.

Figure 1
Figure 1. Overview of Sen-Cap. With sequential point clouds from movable LiDARs and 2D key points from movable cameras as input, Sen-Cap consists of two primary modules, a Unified Across-Sensor Motion Estimator (UAME) for obtaining the human global orientation, pose, and shape, and a Noise-resistant Trajectory Tracker (NTT) for global trajectory estimation. All components corresponding to our key contributions are highlighted… view at source ↗
Figure 2
Figure 2. Systematic drift arises when the true root joint (green ball) is offset from the point cloud center xpc (gray ball), espe￾cially under severe occlusion and noise. Besides a precise local motion, an accurate global trajectory is another critical component for coherent hu￾man motion capture. Thus, we then take as input the sequential point clouds xpc from each LiDAR, and the corresponding 3D joints xˆ L 3d (Sec￾tion 3… view at source ↗
Figure 3
Figure 3. Qualitative comparisons showing Sen-Cap’s superior motion capture in chal￾lenging scenarios: (1) sparse-point leg stretching (2) Outdoor cycling (3) Fast-motion Taekwondo. We have highlighted the regions with significant differences using purple boxes. We also report a LiDAR-only variant (Ours† ) using a single LiDAR sensor. It achieves performance comparable to LiDAR-only methods, while the full multi￾modal version… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Application on a basketball example to illustrate the versatility of Sen-Cap. The results (2) confirm that Sen-Cap successfully fuses heterogeneous inputs from a camera (3) and a LiDAR sensor (1), enabling accurate 3D human pose estimation in challenging dynamic scenar…
Figure 5
Figure 5. Figure 5: Sensor flexibility on FreeMotion-indoor. We visualize Joint Err(PS) for (a) scaling down from a unified 3C+3L model, (b) scaling up a sparse 1C+1L model with additional test-time sensors, and (c) single-modal LiDAR-only and camera-only configurations. Lower values are …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 2 canonical work pages

  1. [1]

    In: GCPR

    Alldieck, T., Kassubeck, M., Wandt, B., Rosenhahn, B., Magnor, M.: Optical flow- based 3d human motion estimation from monocular video. In: GCPR. pp. 347–360. Springer (2017)

  2. [2]

    In: 2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2017)

    Belagiannis, V., Zisserman, A.: Recurrent human pose estimation. In: 2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2017). pp. 468–475 (2017).https://doi.org/10.1109/FG.2017.64

  3. [3]

    Pervasive and Mobile Computing14, Pages–414 (2017)

    Betancourt, A., Díaz-Rodríguez, N., Barakova, E., Marcenaro, L., Rauterberg, M., Regazzoni, C.: Unsupervised understanding of location and illumination changes in egocentric videos. Pervasive and Mobile Computing14, Pages–414 (2017)

  4. [4]

    arXiv preprint arXiv:2308.14492 (2023)

    Cai, Z., Pan, L., Wei, C., Yin, W., Hong, F., Zhang, M., Loy, C.C., Yang, L., Liu, Z.: Pointhps: Cascaded 3d human pose and shape estimation from point clouds. arXiv preprint arXiv:2308.14492 (2023)

  5. [5]

    In: CVPR

    Cao, C., Zhang, Y., Wu, Y., Lu, H., Cheng, J.: Egocentric gesture recognition using recurrent 3d convolutional neural networks with spatiotemporal transformer modules. In: CVPR. pp. 3763–3771 (2017)

  6. [6]

    In: CVPR

    Charles, J., Pfister, T., Magee, D., Hogg, D., Zisserman, A.: Personalizing human video pose estimation. In: CVPR. pp. 3063–3072 (2016)

  7. [7]

    ACM Trans

    Dou, M., Davidson, P., Fanello, S.R., Khamis, S., Kowdle, A., Rhemann, C., Tankovich, V., Izadi, S.: Motion2fusion: Real-time volumetric performance cap- ture. ACM Trans. Graph.36(6), 246:1–246:16 (Nov 2017)

  8. [8]

    mpi - inf

    EgoCap: EgoCap dataset.http : / / gvv . mpi - inf . mpg . de / projects / EgoCap/ (2016), accessed 28 Jun 2026

Show all 57 references
  1. [9]

    arXiv preprint arXiv:2308.00628 (2023)

    Fan, B., Wang, S., Guo, W., Zheng, W., Feng, J., Zhou, J.: Human-m3: A multi- view multi-modal dataset for 3d human pose estimation in outdoor scenes. arXiv preprint arXiv:2308.00628 (2023)

  2. [10]

    arXiv preprint arXiv:2311.11971 (2023)

    Fan, B., Zheng, W., Feng, J., Zhou, J.: Lidar-hmr: 3d human mesh recovery from lidar. arXiv preprint arXiv:2311.11971 (2023)

  3. [11]

    In: International Confer- ence on Computer Vision (ICCV) (2023)

    Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa*, A., Malik*, J.: Humans in 4D: Reconstructing and tracking humans with transformers. In: International Confer- ence on Computer Vision (ICCV) (2023)

  4. [12]

    ACM Transactions on Graphics (TOG)37(6), 1–15 (2018)

    Huang, Y., Kaufmann, M., Aksan, E., Black, M.J., Hilliges, O., Pons-Moll, G.: Deep inertial poser: Learning to reconstruct human pose from sparse inertial mea- surements in real time. ACM Transactions on Graphics (TOG)37(6), 1–15 (2018)

  5. [13]

    In: ECCV (September 2018)

    Huang, Z., Li, T., Chen, W., Zhao, Y., Xing, J., LeGendre, C., Luo, L., Ma, C., Li, H.: Deep volumetric video from very sparse multi-view performance capture. In: ECCV (September 2018)

  6. [14]

    arXiv preprint arXiv:2309.09314 (2023)

    Jang, D.K., Yang, D., Jang, D.Y., Choi, B., Jin, T., Lee, S.H.: Movin: Real-time motion capture using a single lidar. arXiv preprint arXiv:2309.09314 (2023)

  7. [15]

    In: 2017 IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR)

    Jiang, H., Grauman, K.: Seeing invisible poses: Estimating 3d body pose from ego- centric video. In: 2017 IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR). pp. 3501–3509 (2017).https://doi.org/10.1109/CVPR.2017.373

  8. [16]

    arXiv preprint arXiv:2303.07399 (2023)

    Jiang, T., Lu, P., Zhang, L., Ma, N., Han, R., Lyu, C., Li, Y., Chen, K.: Rtm- pose: Real-time multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399 (2023)

  9. [17]

    In: CVPR (2018)

    Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: CVPR (2018)

  10. [18]

    In: ICCV

    Kaufmann, M., Zhao, Y., Tang, C., Tao, L., Twigg, C., Song, J., Wang, R., Hilliges, O.: Em-pose: 3d human pose estimation from sparse electromagnetic trackers. In: ICCV. pp. 11510–11520 (2021) Sen-Cap 17

  11. [19]

    IRAL4(2), 1940– 1947 (2019)

    Kim, W., Ramanagopal, M.S., Barto, C., Yu, M.Y., Rosaen, K., Goumas, N., Vasudevan, R., Johnson-Roberson, M.: Pedx: Benchmark dataset for metric 3-d pose estimation of pedestrians in complex urban intersections. IRAL4(2), 1940– 1947 (2019)

  12. [20]

    arXiv preprint arXiv:2203.14698 (2022)

    Li, J., Zhang, J., Wang, Z., Shen, S., Wen, C., Ma, Y., Xu, L., Yu, J., Wang, C.: Lidarcap: Long-range marker-less 3d human motion capture with lidar point clouds. arXiv preprint arXiv:2203.14698 (2022)

  13. [21]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Li,J., Cao,J., Zhang, H., Rempe, D.,Kautz, J., Iqbal, U.,Yuan,Y.: Genmo: Agen- eralist model for human motion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11766–11776 (2025)

  14. [22]

    ACM Trans

    Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM Trans. Graph.34(6), 248:1–248:16 (Oct 2015)

  15. [23]

    ACM Transactions on Graphics, (Proc

    Loper, M.M., Mahmood, N., Black, M.J.: MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia)33(6), 220:1–220:13 (Nov 2014).https://doi.org/10.1145/2661229.2661273

  16. [24]

    In: ICCV (October 2019)

    Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV (October 2019)

  17. [25]

    Malleson, C., Collomosse, J., Hilton, A.: Real-time multi-person motion capture from multi-view video and imus. IJCV pp. 1–18 (2019)

  18. [26]

    In: 3DV (2017).https://doi.org/10.1109/3dv.2017.00064

    Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 3DV (2017).https://doi.org/10.1109/3dv.2017.00064

  19. [27]

    Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 3DV. pp. 506–516. IEEE (2017)

  20. [28]

    OptiTrack Motion Capture Systems.https://www.optitrack.com/(2009), ac- cessed 28 Jun 2026

  21. [29]

    ACM Trans

    Park, S.I., Hodgins, J.K.: Data-driven modeling of skin and muscle deformation. ACM Trans. Graph.27(3), 1–6 (Aug 2008).https://doi.org/10.1145/1360612. 1360695

  22. [30]

    In: CVPR (Jun 2014).https://doi.org/10.1109/CVPR.2014.300

    Pons-Moll, G., Fleet, D.J., Rosenhahn, B.: Posebits for monocular human pose estimation. In: CVPR (Jun 2014).https://doi.org/10.1109/CVPR.2014.300

  23. [31]

    Qiu, H., Wang, C., Wang, J., Wang, N., Zeng, W.: Cross view fusion for 3d human poseestimation.In:2019IEEE/CVFInternationalConferenceonComputerVision (ICCV). pp. 4341–4350 (2019).https://doi.org/10.1109/ICCV.2019.00444

  24. [32]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Rajasegaran, J., Pavlakos, G., Kanazawa, A., Malik, J.: Tracking people by pre- dicting 3d appearance, location and pose. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 2740–2749 (2022)

  25. [33]

    ACM Trans

    Raskar, R., Nii, H., deDecker, B., Hashimoto, Y., Summet, J., Moore, D., Zhao, Y., Westhues, J., Dietz, P., Barnwell, J., Nayar, S., Inami, M., Bekaert, P., Noland, M., Branzoi, V., Bruns, E.: Prakash: lighting aware motion capture using photosensing markers and multiplexed il...

  26. [34]

    ArXivabs/2407.09833 (2024)

    Ren, Y., Han, X., Yao, Y., Long, X., Sun, Y., Ma, Y.: Livehps++: Robust and coherent motion capture in dynamic free environment. ArXivabs/2407.09833 (2024)

  27. [35]

    arXiv preprint arXiv:2402.17171 (2024) 18 Xue and Sun, A

    Ren, Y., Han, X., Zhao, C., Wang, J., Xu, L., Yu, J., Ma, Y.: Livehps: Lidar-based scene-level human pose and shape estimation in free environment. arXiv preprint arXiv:2402.17171 (2024) 18 Xue and Sun, A. et al

  28. [36]

    TVCG (2023)

    Ren, Y., Zhao, C., He, Y., Cong, P., Liang, H., Yu, J., Xu, L., Ma, Y.: Lidar- aid inertial poser: Large-scale human motion capture by sparse inertial and lidar sensors. TVCG (2023)

  29. [37]

    In: SIGGRAPH Asia Conference Proceedings (2024)

    Shen,Z.,Pi,H.,Xia,Y.,Cen,Z.,Peng,S.,Hu,Z.,Bao,H.,Hu,R.,Zhou,X.:World- grounded human motion recovery via gravity-view coordinates. In: SIGGRAPH Asia Conference Proceedings (2024)

  30. [38]

    In: IEEE/CVF Conf

    Shin, S., Kim, J., Halilaj, E., Black, M.J.: WHAM: Reconstructing world-grounded humans with accurate 3D motion. In: IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (Jun 2024)

  31. [39]

    PLOS ONE11(3) (03 2016)

    Song, M.H., Godøy, R.I.: How fast is your body motion? determining a sufficient frame rate for an optical motion tracking system using passive markers. PLOS ONE11(3) (03 2016)

  32. [40]

    In: IEEE/CVF Conf

    Sun, Y., Bao, Q., Liu, W., Mei, T., Black, M.J.: TRACE: 5D Temporal Regression of Avatars with Dynamic Cameras in 3D Environments. In: IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (June 2023)

  33. [41]

    In: Proceedings of the computer vision and pattern recognition conference

    Wang, Y., Sun, Y., Patel, P., Daniilidis, K., Black, M.J., Kocabas, M.: Prompthmr: Promptable human mesh recovery. In: Proceedings of the computer vision and pattern recognition conference. pp. 1148–1159 (2025)

  34. [42]

    arXiv preprint arXiv:2403.17346 (2024)

    Wang, Y., Wang, Z., Liu, L., Daniilidis, K.: Tram: Global trajectory and motion of 3d humans from in-the-wild videos. arXiv preprint arXiv:2403.17346 (2024)

  35. [43]

    IEEE Sensors Journal 24(9), 15531–15539 (2024).https://doi.org/10.1109/JSEN.2024.3373192

    Wu, X., Zhang, H., Kong, C., Wang, Y., Ju, Y., Zhao, C.: Lidar-based 3-d human pose estimation and action recognition for medical scenes. IEEE Sensors Journal 24(9), 15531–15539 (2024).https://doi.org/10.1109/JSEN.2024.3373192

  36. [44]

    In: 2024 IEEE International Conference on Robotics and Au- tomation (ICRA)

    Xin, S., Zhang, Z., Wang, M., Hou, X., Guo, Y., Kang, X., Liu, L., Liu, Y.: Multi- modal 3d human tracking for robots in complex environment with siamese point- video transformer. In: 2024 IEEE International Conference on Robotics and Au- tomation (ICRA). pp. 337–344 (2024).ht...

  37. [45]

    TPAMI pp

    Xu, L., Su, Z., Han, L., Yu, T., Liu, Y., FANG, L.: Unstructuredfusion: Realtime 4d geometry and texture reconstruction using commercialrgbd cameras. TPAMI pp. 1–1 (2019)

  38. [46]

    In: CVPR (June 2020)

    Xu, L., Xu, W., Golyanik, V., Habermann, M., Fang, L., Theobalt, C.: Eventcap: Monocular 3d capture of high-speed human motions using an event camera. In: CVPR (June 2020)

  39. [47]

    In: 2018 3rd International Conference on Advanced Robotics and Mechatronics (ICARM)

    Xu, W., Li, X., Xu, W., Gong, L., Huang, Y., Zhao, Z., Zhao, L., Chen, B., Yang, H., Cao, L., Liu, C.: Human-robot interaction oriented human-in-the-loop real-time motion imitation on a humanoid tri-co robot. In: 2018 3rd International Conference on Advanced Robotics and Mecha...

  40. [48]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Xu, Y., Cong, P., Yao, Y., Chen, R., Hou, Y., Zhu, X., He, X., Yu, J., Ma, Y.: Human-centric scene understanding for 3d large-scale scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20349–20359 (2023)

  41. [49]

    In: Advances in Neural Information Processing Systems (2022)

    Xu,Y.,Zhang,J.,Zhang,Q.,Tao,D.:ViTPose:Simplevisiontransformerbaselines for human pose estimation. In: Advances in Neural Information Processing Systems (2022)

  42. [50]

    v39i9.32977 Sen-Cap 19

    Xue, A., Ren,Y., Song,Z., Ye, M., Zhu, X.,Ma, Y.:Freecap:hybridcalibration-free motion capture in open environments (2025).https://doi.org/10.1609/aaai. v39i9.32977 Sen-Cap 19

  43. [51]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yan, M., Zhang, Y., Cai, S., Fan, S., Lin, X., Dai, Y., Shen, S., Wen, C., Xu, L., Ma, Y., Wang, C.: Reli11d: A comprehensive multimodal human motion dataset and method. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2250–2262...

  44. [52]

    In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)

    Ye, V., Pavlakos, G., Malik, J., Kanazawa, A.: Decoupling human and camera motion from videos in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)

  45. [53]

    In: CVPR (June 2022)

    Yi, X., Zhou, Y., Habermann, M., Shimada, S., Golyanik, V., Theobalt, C., Xu, F.: Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In: CVPR (June 2022)

  46. [54]

    ACM Transactions on Graphics (TOG)40(4), 1–13 (2021)

    Yi, X., Zhou, Y., Xu, F.: Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions on Graphics (TOG)40(4), 1–13 (2021)

  47. [55]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    Yuan, Y., Iqbal, U., Molchanov, P., Kitani, K., Kautz, J.: Glamr: Global occlusion- aware human mesh recovery with dynamic cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

  48. [56]

    In: ECCV (Sept 2018)

    Zheng, Z., Yu, T., Li, H., Guo, K., Dai, Q., Fang, L., Liu, Y.: Hybridfusion: Real- time performance capture using a single depth sensor and sparse imus. In: ECCV (Sept 2018)

  49. [57]

    2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation rep- resentations in neural networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5738–5746 (2018)

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.