acceptodds
Under review as a conference paper at ICLR 2027

MISP: Mamba-Based Inertial Shape and Pose Estimation with Ultra-Wideband Ranging and Skeletal Priors

Abstract

Sparse inertial pose estimation (SIP) recovers full-body motion from body-worn inertial measurement units (IMUs), offering an occlusion-robust and portable solution for motion capture. Yet under sparse sensor layouts(e.g., 6 IMUs), the problem remains severely under-constrained and suffers from IMU drift. Recent work adds new sensing modalities (e.g., ultra-wideband [UWB] ranging) to mitigate IMU drift. However, prior work is constrained by two main challenges: (1) lack of effective means to acquire and exploit subject-specific body constraints, degrading cross-subject generalization and leaving ranging cues scale-ambiguous; (2) how to effectively exploit heterogeneous observations in temporal modeling. To address these challenges, we propose: (1) MISP, a sparse inertial framework that fuses both UWB-measured inter-sensor distances and subject-specific bone-length priors into the inertial pipeline; to obtain the skeletal prior, MISP-Shaper independently estimates explicit bone-length constraints from a short natural-motion window; (2) Continuous Unified Fusion (CUF), which repeatedly re-injects observation features across temporal blocks to preserve inertial and geometric cues; and Together, these components enable MISP to reduce joint position error by up to 12% over prior work.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
human pose estimation, motion capture, inertial sensors, sensor fusion, state space models, time series modeling, IMU

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_mrGQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents a motion reconstruction method using Mamba blocks to process IMU and UWB sensors. IMU-based mocap suffers from drift due to its inertial nature, therefore, the authors use UWB sensors to compute the distance between sensors and correct the drift from the IMUs. The main contributions are the prediction of the user shape (body segments and SMPL shape parameters), and the Continuous Unified Fusion (CUF) module. The paper is evaluated in two mocap datasets (with synthetic IMU and UWB data) and one dataset with real IMU and UWB data.

Critical strengths
  1. The topic addressed in this paper is interesting and of relevance to obtain consumer-grade reliable mocap. Fusing UWB sensors with IMUs is a promising research direction as demonstrated by previous research.
Critical weaknesses
  1. Limited technical novelty:

The proposed approach is very similar to existing literature in IMU motion reconstruction, and some design choices are similar to previous works already using IMU+UWB sensors. There is little new information to extract from this paper. The pose representation and the architectural details are quite standard, and no new data is presented.

Additionally, some technical details are overlooked and not clearly explained, for example, why the input to the system is orientations instead of the raw information of IMUs which is angular velocities? What is the physics optimizer?

What is the main difference between this method and UIP? For example, one of the main claims is adding shape information, however, in the ablation we can see that it has a marginal effect in the final accuracy (SIP from 9.07 to 9.04). Only when distances are removed there is a clear degradation of performance. However, including distances extracted from UWB is not the key novelty of this paper, as many related works has already proposed the use of these sensors.

  1. Weak Evaluation:

The quantitative evaluation left me with several questions. First of all, there is little difference between MISP and other methods such as UDP in Table 1. Similarly, in Figure 3 there is almost no difference between MISP and UIP.

Why some methods are only shown in some experiments? For example, ASIP is in Figure 3 but not in Table 1, UDP is not shown in Figure 3, the qualitative companion video only compares with ASIP and PNP.

The qualitative comparison only shows basic locomotion, more challenging motions could showcase the benefits of the method more clearly.

In Line 363, what is an acceptable level of jitter? Minimizing jitter is not always desirable as it can produce over smoothing. Provide the jitter value of the dataset.

Why the standalone shape accuracy is not evaluated in the real data on the UIP dataset?

Questions for authors

What is a clear difference in your proposed approach compared to previous work like UIP? If any, which evaluation clearly shows the importance of this difference?

Lines 185–186: Is UWB truly drift-free? Does it require external devices or a fixed reference frame to achieve drift-free tracking? As I understand it, UWB primarily measures distances between devices and does not inherently provide a global reference frame. One of the main challenges of IMU-based methods is root drift, and it is unclear to me how UWB alone can address this issue. Could the authors clarify how root drift is handled in their approach?

2 comments
AuthorsOpenReview ↗

Response to Reviewer mrGQ

TL;DR:

  1. UWB & root drift (Q2): Clarified in Sec. 3.1 & 3.5 that pairwise UWB distances only bound local articulation drift (Θ^t\hat{\boldsymbol{\Theta}}_t). Global root translation pt\mathbf{p}_t is recovered via a shape-scaled contact-physics optimizer, reducing cumulative drift over distance (Fig. 9b).
  2. Distinction from UIP & shape gain (Q1, W1, W3): MISP pairs explicit bone lengths b\mathbf{b} with depth-adaptive CUF, cutting Pos Err over UIP by 17.1–46.6%17.1\text{--}46.6\% across all datasets (Tables 1–2). Skipping opening T-poses via our motion gate lowers TotalCapture SIP to 8.77∘8.77^\circ (Tables 1 & 5).
  3. New evaluations (W2, W4): Added real-UIP standalone shape eval (0.91 cm0.91\text{ cm} BLE, Table 3), dataset GT jitter floors (0.350.35 on TC, 1.251.25 on DIP-IMU, App. A.7), and clarified input/optimizer/video details.

Q2 (Is UWB drift-free & how is global root drift handled?): We agree completely: inter-sensor UWB distances dt∈R15\mathbf{d}_t \in \mathbb{R}^{15} measure internal pairwise geometry and provide no external global reference. As clarified in Sec. 3.1 (L159) and Sec. 3.5 (L291), UWB is "drift-free" only in bounding local pelvis-relative articulation (Θ^t\hat{\boldsymbol{\Theta}}_t) without integrating inertial history. Global translation pt∈R3\mathbf{p}_t \in \mathbb{R}^3 is recovered downstream using PNP's contact-driven physics optimizer (Yi et al., 2024) on a kinematic chain scaled by b^\hat{\mathbf{b}}, using FK(Θ^t;b^)\text{FK}(\hat{\boldsymbol{\Theta}}_t; \hat{\mathbf{b}}) and foot-contact constraints. In Fig. 9b (p. 23), cumulative translation drift still grows with distance, but better local poses and leg lengths give MISP lower drift than PNP and SAIP.

Q1 & W1, W3 (Distinction from UIP & why shape matters):

  • UIP comparison: UIP uses a fixed template body, and UMotion/SAIP regress entangled SMPL β\boldsymbol{\beta}. On real UIP hardware (Table 3, p. 8), soft-tissue shifts inflate β\boldsymbol{\beta} MAE (1.471.47) and surface Mesh-T (3.78 cm3.78\text{ cm}), whereas our explicit bone lengths stay sub-centimetre (0.91 cm0.91\text{ cm} BLE). Paired with CUF's depth-wise routing (Fig. 3b, Table 7), MISP cuts Pos Err over UIP from 5.495.49 to 3.41 cm3.41\text{ cm} on TotalCapture, 5.365.36 to 2.86 cm2.86\text{ cm} on DIP-IMU (Table 1), and 10.6510.65 to 8.83 cm8.83\text{ cm} (−17.1%-17.1\%) on real UIP data (Table 2).
  • Why w/o shape was 9.07∘→9.04∘9.07^\circ \to 9.04^\circ (Sec. 4.4.2, Tables 4–5): w/o shape in Table 4 was retrained from scratch (absorbing average scale); on our frozen estimator (Table 5, p. 10), the Average template gives 11.41∘11.41^\circ SIP (4.76 cm4.76\text{ cm} Pos Err). Moreover, 32/45 TotalCapture sequences open with a motionless T-pose; skipping static openings via our unsupervised motion gate (Eq. 5, p. 17) lowers SIP to 8.77∘8.77^\circ (3.41 cm3.41\text{ cm} Pos Err), closing 72%72\% of the gap from Anatomy only (9.05∘9.05^\circ) to GT bones (8.66∘8.66^\circ).

W2 & W4 (Technical & evaluation details):

  • Inputs & optimizer (Sec. 3.1 L141, Sec. 3.5 L293): Following DIP/PIP/PNP/UIP, IMU nodes output 60 Hz root-relative orientations Rt\mathbf{R}_t via high-rate onboard fusion filters (avoiding 60 Hz gyro integration error), and global translation uses PNP's contact-physics solver scaled by b^\hat{\mathbf{b}}.
  • Real UIP shape, GT jitter & video (Table 3, App. A.7 & A.9, p. 20–23): Zero-shot BLE on real UIP streams is 0.91 cm0.91\text{ cm} (Table 3). Ground-truth motion jerk floors are 0.350.35 (TC) and 1.251.25 (DIP-IMU); MISP achieves 0.280.28 (0.180.18 with w/ dynopt) and 0.140.14, staying below GT jerk and beating ASIP in 1–13 Hz PSD (Fig. 8, where full-sequence SIP is 8.10∘8.10^\circ vs. 8.23∘8.23^\circ). The video compares ASIP (window seams) and PNP (contact-pinned stiffness) as opposite smoothness baselines and includes dynamic turning and freestyle motions (Fig. 9a).

Thank you for your response. I understand much better some of my concerns related to clarity of exposition and soundness. However, the main weaknesses of the paper about novelty and evaluation are still persistent and would require a major revision. Therefore I am keeping my original score.

Reviewer_S2PvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies causal full-body motion estimation from six body-worn IMU–UWB sensor nodes. MISP combines inertial measurements, inter-sensor distances, and a subject-specific skeletal prior in a Mamba-2 pose estimator. A separate Shaper estimates bone lengths and body shape from a short natural-motion IMU window conditioned on anthropometric measurements. The resulting skeletal prior is held fixed during subsequent tracking. Continuous Unified Fusion (CUF) supplies layer-specific projections of the current fused observation to successive temporal blocks. A separate physics optimizer handles global translation. Experiments on TotalCapture, DIP-IMU, and UIP examine pose accuracy, shape estimation, component contributions, and efficiency. The reported results show competitive positional accuracy, alongside accuracy–smoothness trade-offs.

Critical strengths

In the submitted Table 1, MISP achieves positional errors of 3.44 cm on TotalCapture and 2.87 cm on DIP-IMU, the lowest among the listed methods.

Critical weaknesses

1.Table 1 lists UMotion's TotalCapture angular error as 11.36 degrees, whereas its original Table 1 reports 7.06 degrees. Different joint sets or re-evaluation protocols may explain this discrepancy, but the manuscript does not reconcile it. Section 4.1 also reports fine-tuning on UIP and DIP-IMU without making the corresponding treatment of all baselines clear. 2.The Shaper is a central contribution, but the reported results do not isolate how much natural-motion IMU evidence adds beyond anthropometric measurements alone. The no-shape ablation changes SIP by only 0.03 degrees while worsening positional error by 0.18 cm; the average/predicted/ground-truth prior comparison shows a much larger SIP difference. These results need not conflict, but their interpretation depends on whether the pose network is retrained or kept fixed. 3.The manuscript contains several writing and consistency issues that should be corrected: (i) the abstract contains the editorial error “and Together” (p. 1, line 27); (ii) the references list both the 2015 preprint and the 2016 conference version of the same ResNet work (p. 10); (iii) Table 2 defines Δ as RANSAC − Top-k, but reports −0.02 for both rows, whereas the listed values yield +0.02 (p. 14). Please correct the wording, consolidate or clarify the duplicate references, and reconcile the definition and signs of Δ. Although these issues do not change the reported method ranking, they reduce the precision and consistency of the presentation.

Questions for authors

1.Could you identify which Table 1 entries were reproduced versus taken from prior publications, and reconcile the UMotion angular-error discrepancy? Please specify the evaluated joints, preprocessing, training-data exclusions, and fine-tuning data for the compared methods. Does the angular-accuracy advantage persist under a common protocol? 2.For the no-shape and average/predicted/ground-truth prior experiments, which network parameters are retrained and which are held fixed? Under matched training and evaluation conditions, how does an anthropometrics-only Shaper compare with the full Shaper in both bone-length error and downstream positional error? 3.On DIP-IMU, adding CUF to UMotion improves SIP error from 14.19 to 11.88 degrees but worsens positional error from 3.38 to 4.59 cm and angular error from 6.35 to 6.75 degrees. Are these variants trained and evaluated under identical conditions? How should this trade-off qualify the claim that CUF generalizes beyond MISP?

1 comment
AuthorsOpenReview ↗

Response to Reviewer S2Pv

TL;DR:

  1. UMotion protocol resolved (Q1 & Q3): UMotion's script overwrites the 8 unmeasured distal joints with ground-truth root-relative rotations (zeroing out error on 8/248/24 joints), whereas DIP, PNP, and our pipeline set their local rotations to identity so parent errors propagate via FK (App. A.5). Evaluated under the unified protocol in Table 1, UMotion+CUF improves all metrics on both datasets, and MISP leads overall.
  2. Zero-shot on DIP-IMU & TotalCapture (Q1): All Table 1 models are trained only on AMASS; only UIP (Table 2) is fine-tuned on its training split (Sec. 4.1).
  3. Shaper isolated (Q2): Clarified retrained w/o shape (Table 4) vs. frozen-estimator swaps (Table 5), and added a motion gate (Eq. 5) that improves TotalCapture SIP from 9.05∘9.05^\circ (Anatomy only) to 8.77∘8.77^\circ (Full gated, vs. 8.66∘8.66^\circ GT oracle).
  4. Typos fixed: Fixed L28 typo, duplicate ResNet citation, and removed the inverted-sign Δ\Delta column in Table 3.

Q1 & Q3 (Table 1 fairness, UMotion discrepancy & fine-tuning splits):

  • 8-unmeasured-joint convention (App. A.5, p. 16): Eight distal joints (ankles, feet, wrists, hands) have no IMUs. (1) Standard local-identity FK protocol (DIP, PNP, ours): sets their parent-relative (local) rotations to identity before FK, so parent orientation errors (knees, elbows, etc.) propagate along the chain to all 24 joints in root-relative space. (2) UMotion's official script: operates in root-relative space and overwrites those 8 joints with ground-truth rotations, giving 8/248/24 joints (one-third of the body) strictly 0∘0^\circ error and scaling down the 24-joint mean by ∼1/3\sim 1/3 (11.36∘×16/24≈7.57∘11.36^\circ \times 16/24 \approx 7.57^\circ vs. 7.06∘7.06^\circ). Initially, we re-evaluated UMotion on TotalCapture with local-identity FK (11.36∘11.36^\circ), but inadvertently copied UMotion's GT-filled paper numbers on DIP-IMU (6.35∘/3.38 cm6.35^\circ / 3.38\text{ cm}) while scoring UMotion+CUF with local-identity FK (6.75∘/4.59 cm6.75^\circ / 4.59\text{ cm}). In revised Table 1 (p. 6), all methods use the unified local-identity FK protocol: on DIP-IMU, UMotion scores 14.19∘14.19^\circ SIP, 7.46∘7.46^\circ Ang Err, 4.86 cm4.86\text{ cm} Pos Err, and UMotion+CUF improves all three to 11.88∘11.88^\circ (−16.3%-16.3\%), 6.75∘6.75^\circ (−9.5%-9.5\%), and 4.59 cm4.59\text{ cm} (−5.6%-5.6\%) (and similarly on TotalCapture: 10.76∘→9.60∘10.76^\circ \to 9.60^\circ, 11.36∘→10.42∘11.36^\circ \to 10.42^\circ, 4.46→3.48 cm4.46 \to 3.48\text{ cm}). There is no trade-off, and MISP leads across all metrics (9.65∘/3.41 cm9.65^\circ / 3.41\text{ cm} on TC; 5.72∘/2.86 cm5.72^\circ / 2.86\text{ cm} on DIP-IMU).
  • Training splits (Sec. 4.1 L323): No fine-tuning is done on TotalCapture or DIP-IMU (the DIP-IMU train split only fits the two scalar calibration constants in Eq. 3). Only UIP (Table 2) is fine-tuned on its training split, matching all baselines.

Q2 (Retrained vs. frozen Shaper ablations): In Table 4, w/o shape (9.07∘9.07^\circ SIP, 3.62 cm3.62\text{ cm} Pos Err) is retrained from scratch without bone channels (absorbing average scale into weights, though still trailing gated MISP at 8.77∘/3.41 cm8.77^\circ / 3.41\text{ cm}). In Table 5 (p. 10), the pose estimator is frozen (Average template gives 11.41∘/4.76 cm11.41^\circ / 4.76\text{ cm}). Skipping motionless opening T-poses on TotalCapture (32/45 sequences) via our unsupervised motion gate (Eq. 5, App. A.6) improves BLE from 0.660.66 to 0.58 cm0.58\text{ cm} on TC and 0.360.36 to 0.33 cm0.33\text{ cm} on DIP-IMU (Full vs. Anatomy, Table 3), and lowers downstream pose error from 9.05∘/3.52 cm9.05^\circ / 3.52\text{ cm} (Anatomy only) to 8.77∘/3.41 cm8.77^\circ / 3.41\text{ cm} (Predicted gated, Table 5), closing 72%72\% of the gap to GT bones (8.66∘/3.30 cm8.66^\circ / 3.30\text{ cm}).

Minor Fixes: Fixed L28 typo, duplicate ResNet entry (L588), and removed the redundant Δ\Delta column in Table 3 whose sign was inverted in the initial draft.

Reviewer_3kgkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MISP combines sparse IMU measurements, UWB distances, and estimated bone lengths for human pose estimation. A separate Shaper provides a static skeletal prior, while a Mamba-2 network uses layer-wise observation reinjection (CUF) to predict pose. Experiments on TotalCapture, DIP-IMU, and UIP demonstrate competitive position and angular accuracy.

Critical strengths

The paper is clearly written and presents a coherent pipeline separating static shape estimation from dynamic pose tracking. Combining inertial and geometric information is well motivated, and results across three datasets support the effectiveness of the overall approach.

Critical weaknesses
  1. The methodological contribution beyond the existing ingredients remains insufficiently established. The paper explains its design differences from prior work, including explicit skeletal priors and layer-specific observation reinjection. However, the experiments mainly show component-wise improvements and provide limited evidence of benefits specific to their integration. The connection and depth ablations support CUF’s effectiveness, but do not isolate observation dilution as the mechanism behind its gains.
  2. Important contributions remain difficult to attribute. Mamba-2’s role in the overall gains is not isolated through a controlled backbone comparison. Separately, Shaper lacks an input ablation separating the contribution of motion observations from anthropometric information. The reported bone-length accuracy and downstream prior comparisons do not establish how much subject-specific information is recovered from motion.
Questions for authors
  1. Do CUF’s benefits change when skeletal or ranging inputs are removed? Is there evidence supporting the observation-dilution explanation beyond the reported accuracy and jitter changes?
  2. How does Mamba-2 compare with another causal backbone when inputs, CUF, model capacity, and training conditions are comparable?
  3. How much does motion input improve Shaper over anthropometrics alone? Beyond the RANSAC comparison, does calibrated top-k outperform simple averaging? Please also clarify whether the no-shape and average-bone variants are retrained or modified only at inference.
  4. Beyond the UIP evaluation, is there evidence from live IMU/UWB acquisition, including sensitivity to ranging bias or sensor-placement variation? If live tests were conducted, what end-to-end latency was measured, including communication and preprocessing?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 3kgk

TL;DR:

  1. CUF (Q1): Dropping CUF costs 4.2%4.2\% SIP without UWB vs. 12.4%12.4\% with hybrid inputs (Table 4); layer attributions confirm distance-to-orientation routing across depth (Fig. 3b, Fig. 6, Table 7).
  2. Backbone (Q2): Mamba-2 beats a matched 3-layer LSTM by 4.9%4.9\% SIP (9.51∘→9.04∘9.51^\circ \to 9.04^\circ, Table 4) with a 2.3×2.3\times online / >20×>20\times offline GPU speedup (Table 9).
  3. Shaper (Q3): Skipping opening T-poses via our motion gate (Eq. 5) lowers TotalCapture SIP to 8.77∘8.77^\circ, closing 72%72\% of the gap from Anatomy only (9.05∘9.05^\circ) to GT bones (8.66∘8.66^\circ, Table 5).
  4. Hardware & latency (Q4): Added real-UIP shape eval (0.91 cm0.91\text{ cm} BLE, Table 3), 30%30\% UWB dropout (+0.09∘+0.09^\circ SIP, Table 8), and latency benchmarks (Table 9).

Q1 (CUF mechanism): Removing CUF on an IMU-only stack (w/o distance, w/o CUF vs. w/o distance, Table 4, App. A.6 L878) degrades SIP by 4.2%4.2\% (12.31∘→12.83∘12.31^\circ \to 12.83^\circ), vs. 12.4%12.4\% (9.04∘→10.16∘9.04^\circ \to 10.16^\circ) on the hybrid stack, showing CUF's benefit scales with input heterogeneity. Probing ψ(l)∘ϕin\psi^{(l)} \circ \phi_{\text{in}} (Fig. 3b, Table 7, p. 19) shows Layer 1 focuses on UWB distances (63.4%63.4\% weight) to bound the cross-limb envelope, while Layer 3 shifts to IMU orientations (68.0%68.0\%) to resolve axial twist. Additive residuals (w/ residual) distort ϕin\phi_{\text{in}}, spiking acceleration weight 3.5×3.5\times and cutting bone-prior weight by 6.2%6.2\% (Fig. 6).

Q2 (Backbone): Under identical inputs, CUF, shaper prior, and 270-epoch schedule, Mamba-2 outperforms a matched-width 3-layer LSTM (w/ LSTM, Table 4, App. A.6 L874): 9.04∘9.04^\circ vs. 9.51∘9.51^\circ SIP and 3.443.44 vs. 3.46 cm3.46\text{ cm} Pos Err, while running 2.3×2.3\times faster online (3,876 FPS, Table 9).

Q3 (Shaper):

  • Inputs & motion gate (Table 3, Table 5, App. A.6 p. 17–18): Full (0.33/0.58 cm0.33 / 0.58\text{ cm} BLE on DIP/TC) beats Anatomy (0.36/0.66 cm0.36 / 0.66\text{ cm}) and IMU only (0.48/0.65 cm0.48 / 0.65\text{ cm}). Because 32/45 TotalCapture sequences start with a motionless T-pose, gating the 1.5 s window on limb orientation spread (Eq. 5) lowers SIP from 9.04∘9.04^\circ to 8.77∘8.77^\circ (3.41 cm3.41\text{ cm} Pos Err, Table 1 & 5), closing 72%72\% of the gap from Anatomy only (9.05∘9.05^\circ) to GT bones (8.66∘8.66^\circ). Uniform scale α\alpha leaves joint angles invariant, whereas distal limb proportions carry 72.2%72.2\% of angular sensitivity and predict ΔSIP\Delta\text{SIP} with R2=0.90–0.94R^2 = 0.90\text{--}0.94 (App. A.6 p. 18).
  • Top-kk & protocol: Power-law recalibration (Eq. 3, Fig. 4, p. 14) avoids the frames 0–25 warmup transient (>1.0 cm>1.0\text{ cm}) that hurts simple window averaging. w/o shape in Table 4 is retrained from scratch; Table 5 uses a frozen estimator (Sec. 4.4 L436).

Q4 (Real hardware & latency): On real UIP hardware, we reach 0.91 cm0.91\text{ cm} BLE zero-shot (Table 3) and cut Pos Err by 14.4–17.1%14.4\text{--}17.1\% (Table 2). While no new live captures were recorded, UIP contains real UWB biases (RANSAC-calibrated, Sec. 4.1) and placement shifts (partly absorbed into b^\hat{\mathbf{b}}; systematic placement testing is left to future work). A 30%30\% UWB dropout raises SIP by only 0.09∘0.09^\circ (Table 8, p. 20). Runtime is 3,876 FPS on GPU (0.26 ms0.26\text{ ms}) and 220 FPS (4.5 ms4.5\text{ ms}) with the CPU physics solver (Table 9, p. 21).

Thank you for the detailed response. The CUF analysis, LSTM comparison, and Shaper ablations address most of my concerns. My remaining concern is how the full system performs in practice. The UIP results are helpful, but they are based on recorded data. It’s still unclear how well the system tracks in actual use, where sensor placement and ranging conditions can vary.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and New Experiments

We thank all reviewers for their constructive feedback. In the revised PDF, we have resolved the evaluation protocol questions and added controlled experiments to isolate each component:

  • CUF Mechanism & Modality Interaction (Reviewer 3kgk): Added an IMU-only CUF ablation (w/o distance, w/o CUF, Table 4) and layer-wise Jacobian/variance attributions (Fig. 3b, Fig. 6, Table 7, App. A.6), showing how CUF transitions from UWB distance bounding (Layer 1) to IMU orientation refinement (Layer 3) while avoiding the input-projection distortion of additive residuals.
  • Controlled Backbone Comparison (Reviewer 3kgk): Added w/ LSTM in Table 4 (App. A.6), comparing Mamba-2 against a matched-width 3-layer unidirectional LSTM under identical inputs, CUF, shaper prior, and training schedule.
  • Shaper Ablations, Motion Gate & Sensitivity (Reviewers 3kgk, S2Pv, mrGQ): Clarified retrained (w/o shape, Table 4) vs. frozen-estimator prior swaps (Table 5); added Full vs. IMU only vs. Anatomy ablations and real-UIP standalone shape evaluation (Table 3); introduced an unsupervised motion gate (Eq. 5, App. A.6) that skips motionless opening T-poses on TotalCapture (9.04∘→8.77∘9.04^\circ \to 8.77^\circ SIP, Table 1); and added a bone-perturbation sensitivity analysis (Sec. 4.4.2, App. A.6, p. 18).
  • Protocol Unification & Jitter Calibration (Reviewers S2Pv, mrGQ): Clarified zero-shot vs. fine-tuning splits (Sec. 4.1), unified the 8-unmeasured-joint evaluation convention across Table 1 (App. A.5), and added ground-truth motion jerk baselines and PSD spectral comparisons (Fig. 8, App. A.7).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.