acceptodds
Under review as a conference paper at ICLR 2027

Cepheid: Decoupling Scale from Structure for Cross-Family Motion Completion

Abstract

Sparse-to-dense motion completion recovers dense mesh motion from sparse vertex displacements. Residual priors trained on one geometry family frequently fail to beat parameter-free harmonic extension off-family. Moreover, cross-family training meets a scaling barrier: residual amplitude varies by more than an order of magnitude across geometry families and collapses as observations become denser. Since this scale depends on unknown target motion, it cannot be measured directly at test time. We decouple residual structure from physical scale by learning a dimensionless prior, estimating amplitude from observables and a lightweight correction fitted only on training sources. Our unified model surpasses harmonic extension across all evaluated families, including unseen cloth and legged bodies, without skeletons, registration, or cross-mesh correspondence. The same prior applies without retraining to point-tracked animal video and a robot grasping a soft body.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
motion completion, flow matching, 3D deformation, sparse observations, cross-domain generalization, residual learning, scale generalization, generative models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bPSVOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents CEPHEID, a unified framework for sparse-to-dense motion completion across different geometry families. The key idea is to decouple residual motion structure from physical scale, combining harmonic extension with a dimensionless learned residual prior and target-independent amplitude recovery. The method generalizes to unseen geometry families without requiring skeletons, registration, or cross-mesh correspondence.

Critical strengths

The problem setting is interesting and practically relevant, especially for cross-family motion completion without skeletons or predefined correspondences.

The motivation for scale decoupling is clear and supported by empirical observations.

The method shows consistent improvements over harmonic extension, including on unseen cloth and legged robot families.

The ablation studies reasonably support the importance of the main components.

Critical weaknesses

The methodological novelty is somewhat incremental, as the framework mainly combines harmonic extension, residual learning, scale normalization, and flow matching.

The baseline coverage for cross-family generalization is relatively limited.

The real-world demonstrations are mostly qualitative.

The quantitative evaluation assumes relatively clean vertex displacement observations, so robustness to realistic tracking noise is not fully demonstrated.

Questions for authors

How does the proposed amplitude recovery compare with simpler per-instance normalization strategies?

Can the authors include comparisons with more recent cross-topology or category-agnostic deformation methods?

How robust is CEPHEID to noisy, missing, or drifting point trajectories in real-world settings?

3 comments

Novelty (W1). Our contribution is a problem and a diagnosis of why learned priors fail on it: completing dense motion from sparse, unlabeled vertex displacements across geometry families, without a skeleton, registration or correspondence. Reviewer 98K5 finds the "problem formulation and motivations are clear", and Reviewer 5qQe that "the problem is useful". In this setting a prior trained on one family transfers poorly, with 7 of the 12 off-diagonal specialist cells in Table 2 at or above harmonic extension, because the residual's amplitude spans 18–39× across families at matched density and depends on the unknown target. As the reviewer notes, "the motivation for scale decoupling is clear and supported by empirical observations". Separating amplitude from the learned structure lets one model trained on DT4D, AMASS and Franka fall below harmonic extension in pooled error in all five columns, including cloth and legged robots that training never saw. The ablation in §5.4 shows that this separation carries the result: without amplitude recovery, transfer turns negative on Franka (0.020 → 0.209, harmonic 0.101) and legged robots (0.223 → 0.409, harmonic 0.247). Harmonic extension, residual learning and flow matching are known components; what is new is identifying amplitude as the obstacle to cross-family transfer and removing it.

Per-instance normalization (Q1). Table 3's "Raw observed scale" row, now a curve in Figure R1, is exactly per-instance normalization by an input statistic: ŝ = σ_obs, as in RevIN-style methods. On cloth it is above harmonic extension at every density. Amplitude recovery is within 1.4–8.1% of the target-informed oracle, which uses the ground-truth target and serves only as a reference. The needed scale is the residual's own dispersion, s = std(F − H), and s/σ_obs is not constant. At the sparsest density, s spans 37.5× across the training sources and s/σ_obs still spans 3.2× (Appendix C.1). We also measured this on the training streams (12,000 conditions). Dividing the residual by σ_obs leaves a 3.1–5.9× spread across sources at each density, and 104× over all source × density cells. With amplitude recovery the spread is 1.1–2.3× per density and 4.1× overall. The raw residual spreads 429×. We also retrained the full model with ŝ = σ_obs, with the same data, backbone and steps, and scored both models on the same conditions (50 pairs per family; all Table 3 conditions on cloth). This experiment is reported here, not in the revised PDF. Ratios are the retrained model's error divided by CEPHEID's, so values above 1 favour CEPHEID. The retrained model is 4.4% worse on held-out cloth (1.044 [1.022, 1.080]) and ties on held-out legged robots (0.988 [0.927, 1.044]). On the training families it is 6.6% worse on AMASS and 15.7% worse on Franka, and 7.1% better on DT4D. Retraining with per-instance normalization therefore does not improve transfer to the unseen families, which is the setting amplitude recovery is designed for: it is worse on cloth and level on legged robots.

  • More baselines and recent cross-topology methods (W2/Q2). See GC1.

  • Real-world demonstrations mostly qualitative (W3). See GC3 (DAVIS horse against held-out tracks, Dynamic FAUST scans) and GC2 (TAPIP3D tracker errors). On the DAVIS horse, CEPHEID is at 4.5 px against 8.7 px for harmonic extension on 51 held-out tracks, and on Dynamic FAUST it is 50.4% below harmonic extension.

  • Noisy, missing or drifting trajectories (W4/Q3). GC2 covers all three: tracker errors that include drift accumulated from frame 0, dropout up to 50%, persistent spatially correlated offsets, identity switches and correlated noise. With TAPIP3D tracks, CEPHEID is 28–30% below harmonic extension on three of four clips and ties on the fourth, and it stays below in all five columns with up to 50% dropout.

Thanks for the detailed response. I found the paper very interesting. But I’m less familiar with this area, so I’m curious about one possible application. If I use a video generator like MiniMax to animate a rigged character, could CEPHEID help recover motion on the character’s original rig from the generated video? If so, could this be a route to text-to-motion generation for characters with arbitrary rigs?

Thank you, we are glad you found the paper interesting. Yes, CEPHEID pairs naturally with video generators such as MiniMax, because it does not require a rig. Rigging is a separate step that still relies heavily on expert work, and automatic rigging of arbitrary shapes is an active research problem; CEPHEID skips it. It completes the motion of every vertex from sparse vertex displacements, using only the mesh connectivity. In the paper, one model handles rigged characters (DT4D animals, AMASS bodies, the Franka arm) and unrigged objects (cloth, unseen in training, and the DAVIS horse from real video, Section 6).

Sparse points obtained from the generated video, for example with a point tracker, are completed into dense motion, which can then be fitted to a rig if one exists. The updated supplementary shows two unrigged characters animated from generated videos, each with about twenty sparse points: an octopus (octopus_generated_video/) and an elephant whose trunk has no bones (elephant_generated_video/). We see this as a promising route to text-to-motion for arbitrary characters.

Reviewer_98K5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a method for dense completion of sparse motions described as vertex displacements without prior correspondence and generalizing across shape categories. It uses harmonic extension of the observed motion and predicts the remaining residual using a conditional flow-matching. For this, it separates residual structure from amplitude.

Critical strengths

Problem formulation and motivations are clear and make sense to me. I agree one of the main limitations with learning-based method for shape deformations and motion synthesis is that it usually cannot generalize across categories.

The proposed method makes a useful base formulation with the harmonic extension which serve as an interesting baseline.

Critical weaknesses

My main concern is about the chosen baselines for comparisons and qualitative results. The provided motion completions suffer from important artifacts (overly stretched triangles in particular). I would expect energy based method (like ARAP or related) to prevent this kind of issue. Speaking of ARAP, this important comparison is missing: if you have sparse vertex trajectories, I would suggest trying to compute iterative ARAP deformations as a non-learning baseline. Additionally, I see lots of references but only two models (NSDP and DOMA) are used for comparisons and

Then, I understand that the paper was written using AI, which is fine, but a substantial effort should be put to make the paper clearer because I find some reported results and interpretation overly crowded. For example, for Table 3, I would suggest using a graph to show the degradation trend.

Finally, about the method itself, I see that it uses a flow-matching network but it is not formulated in the Method section, only mentioned in Figure 2 and in the appendix. This seems important to me, especially, I do not see any discussion on seed variance: it is a generative method so the aim is not really dense completion in the deterministic sense but a generation of plausible intermediate vertex trajectories if I am correct. I see it is indirectly evaluated through averaging in Appendix C.8 but I am not convinced of the method’s stability.

Questions for authors

How much variance do you observe qualitatively when changing seeds ?

How does CEPHEID compare to an ARAP based approach ?

How expensive is the inference in terms of GPU usage and time ?

The method is claimed to not require any correspondence prior: how well does it handle mesh resolution? Does it work on very high resolution scans with millions of vertices ?

5 comments

Baselines beyond NSDP and DOMA. See GC1: TPS, biharmonic interpolation and converged ARAP on all five columns, and DOMA on the two held-out families.

ARAP (Q2). We ran iterative local–global ARAP (Sorkine & Alexa) until the relative change fell below 1e-5, with a cap of 20,000 iterations, raised to 200,000 on the legged robots, where some conditions needed up to 87,338. All conditions converged. Weights are cotangent where faces exist and uniform otherwise, and observed vertices are pinned. Starting from the harmonic or the biharmonic field gives the same converged error to within 1%. CEPHEID's error is 16–29% lower on DT4D, AMASS and Franka, 7% lower on cloth and 27% lower on the legged robots (Table R1). Cloth has one exception. At 0.5% observation, ARAP gets worse as it iterates, and stopping at 10 iterations (0.652) is slightly better than CEPHEID (0.679). We report converged ARAP throughout.

Stretched triangles (Appendix I, Table R5). The stretched triangles visible in Figure 1's animal panel come from drawing, not from the completion. The completion runs on a 3,000-vertex graph and is mapped onto the 20,314-vertex render mesh for the figure. Ground truth drawn through the same mapping stretches as much, with 22.1% of triangles above 1.3× against 23.0% for CEPHEID. When binned by distance to the nearest observed vertex, CEPHEID stretches no more than ground truth in every bin on DT4D, AMASS, Franka and legged robots. Harmonic extension's stretch sits next to the observed vertices. At 0.5% observation in the nearest bin on DT4D, 23.8% of its triangles exceed 1.3×, against 2.7% for CEPHEID. CEPHEID (k = 8) has the lowest mean edge-length error on DT4D, AMASS and legged robots. On DT4D it is 0.024, against 0.030 for ARAP. On cloth ARAP is lower, 0.017 against 0.033, and on DT4D ARAP has fewer edges stretched beyond 1.3×, 0.58% against 0.93%. CEPHEID adds no detectable self-intersections over ground truth (pooled CIs contain 0), while harmonic extension adds 56 per condition on AMASS. On Franka, link lengths change by 0.024% with CEPHEID and 0.045% with ARAP.

Inference cost (Q3; Appendix A.8, Table R2). One completion is a sparse solve, the amplitude rule and 30 Euler steps, with the k = 8 draws batched. On one RTX A6000 it takes 4.5 s (median) and 1.0 GB at 3,000 vertices, 53 s and 6.9 GB at 20k, 235 s and 18.0 GB at 50k, and 778 s and 34.7 GB at 100k. Time grows about as N^1.5 over this range.

Resolution and million-vertex scans (Q4). CEPHEID needs no correspondence, so a high-resolution mesh can be completed on a decimated proxy and mapped back with barycentric coordinates. The network never runs on the full 1M vertices. We measured the cost on a 1M-vertex mesh subdivided from a 3k DT4D mesh, not a captured scan, with one frame pair averaged over three densities. Completion through a 20k proxy takes 57.8 s. Decimation adds about 80 s once per mesh and is shared by all its frames. Native costs up to 100k vertices are in Table R2.

Formulation (W3). §3.4 now gives the objective and the sampler. Let R = F − H be the residual and x₁ = clip(R/ŝ, −8, 8). Training uses the OT path with σ_min = 0.001:

x_t = (1 − (1 − σ_min) t) x₀ + t x₁, x₀ ~ N(0, I); L(θ) = E ‖v_θ(x_t, t | X, Ω, O, H) − (x₁ − (1 − σ_min) x₀)‖² (R1)

At inference we integrate 30 uniform Euler steps for k = 8 draws and output F̂ = H + ŝ · (1/8) Σᵢ x₁⁽ⁱ⁾ (R2).

What the output is. Each condition has one ground-truth motion. The sampler is stochastic, so individual draws differ. We report the mean of k = 8 draws, which estimates the conditional mean that an L2 error rewards. We do not claim sample diversity. A regression model with the same backbone, data and steps (§5.4, Tables 4 and 8) is worse than CEPHEID on cloth (0.415 vs 0.369), Franka (0.028 vs 0.020) and legged (0.226 vs 0.223), and ties on DT4D and AMASS. Beyond the mean, the spread of the 8 draws tells where the completion is uncertain. Specifically, the per-vertex RMS distance of the draws to their mean ranks the 10% of unobserved vertices with the largest error within each condition at a median AUROC of 0.875–0.956 across the five evaluation columns. Graph distance to the nearest observation scores 0.66–0.83 on the same task on DT4D and AMASS. A regression model gives one output and no such signal.

Seed variance (Q1; Appendix C.8, Figure R3). We reran three sampling seeds on 50 pairs per family at all densities. The pooled error of the 8-draw mean changes by 0.2–4% across seeds: DT4D 1.2%, AMASS 0.8%, Franka 4.0%, cloth 0.2%, legged 0.4%. A single draw has 8–21% higher pooled error (Eq. 6) than the mean, with k = 1 ÷ k = 8 at 1.21, 1.20, 1.21, 1.12 and 1.08. Between consecutive frames the 8-draw mean changes about as much as the ground truth does (0.89–1.07×). Single draws jitter at sparse observation, up to 1.6–1.9× the ground truth at 0.5% on DT4D and cloth, which is why we average. They are also worse on most geometric metrics, with 1.4–2.3× the frame-to-frame error (Appendix I). Figure R3 shows the same thing on one median condition per family: the 8-draw means look nearly identical across seeds, while single draws differ, mostly far from the observations. All of these are sampling seeds of one trained model.

Table 3 (W2). Table 3 in §5.3 is now a plot of the trend across densities, printed at full text width as Figure R1. Its numbers remain in Appendix C. To reduce crowding, §4 and §5 move secondary numbers (column denominators, per-density margins, metric factors, baseline settings) to the appendix sections that already cover them. The tightened spacing is also restored. No result or claim changes.

Hi,

Thank you for this response. My concerns are partially addressed. The new table R1 is an unusual way to compare methods making the paper even harder to read. The new comparison to ARAP is good but qualitative examples would be welcome as I am not particularly conviced by the qualitative examples (from supplementary and in figures).

I will maintain ly original score. The paper still has limited comparisons to state-of-the-art approaches. I encourage the authors to revise the evaluation: more qualitative comparisons, a user study could be a strong argument, clearly showing where existing methods fail and what your approach can bring there

Reviewer_5qQeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies dense motion completion from sparse vertex displacements across different geometry families. Cepheid combines harmonic extension with a learned residual and estimates the residual scale from the inputs. It is trained on DT4D, AMASS, and Franka, and evaluated on these sources and unseen cloth and legged robots.

Critical strengths
  1. The problem is useful. One model can handle different geometry families without skeletons or category-specific templates, including families not seen during training.
  2. Separating residual structure from scale is well motivated, and the ablations support the importance of scale recovery.
Critical weaknesses
  1. Comparisons with existing methods on unseen families are missing. NSDP and DOMA are mainly compared on DT4D. Results from TPS and DOMA on cloth and legged robots would help establish the cross-family advantage.
  2. The main experiments use exact ground-truth displacements. There is no controlled noise evaluation, so it is unclear whether the gains hold with imperfect tracking inputs.
  3. The visual benefit over harmonic extension is not clear in several demos. The real horse example is also difficult to assess from static trajectory overlays without a video and multiple rendered views.
Questions for authors
  1. Could you add TPS and DOMA comparisons on the unseen cloth and legged families?
  2. How does the method perform when noise is added to the observed displacements?
  3. Could you provide multi-view visualizations of the results, including the real horse example?
1 comment
  • TPS and DOMA on cloth and legged (W1/Q1). See GC1. On cloth CEPHEID is 0.80× TPS and 0.75× DOMA. On legged it is 0.79× TPS and 0.21× DOMA.

  • Noise (W2/Q2). See GC2: TAPIP3D tracks on rendered videos, dropout, persistent spatially correlated offsets, identity switches and correlated noise on all five families. CEPHEID is 28–30% below harmonic extension on three of four tracker clips and ties on the fourth, and stays below it in all five columns with up to 50% of observations dropped. Appendix H has the complete curves, including i.i.d. noise.

  • Visual benefit, the horse and multi-view (W3/Q3). See GC3. The horse is now quantitative: 4.5 px against 8.7 px on 51 held-out tracks, and a ratio of 0.598 over random input subsets. The supplementary has multi-view videos for all five families and a four-view video of the horse. Visible gains concentrate where observations are sparse and large regions are unobserved. At dense observation harmonic extension is already accurate in absolute terms (Table 1, 50%: 0.075), so the gains there show in the numbers rather than in the renders.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General response (1/2)

We thank the reviewers. Reviewer 5qQe finds that "separating residual structure from scale is well motivated", and Reviewer bPSV that "the motivation for scale decoupling is clear and supported by empirical observations". The submitted results show why the separation matters. A prior trained on one family transfers poorly: 7 of the 12 off-diagonal specialist cells in Table 2 are at or above harmonic extension. The residual F − H differs in amplitude by 18–39× across families, and that amplitude depends on the unknown target. CEPHEID learns the residual's structure in dimensionless form and recovers its scale from the observations. One model trained on DT4D, AMASS and Franka is then below harmonic extension in all five columns, including the held-out cloth and legged robots. Without amplitude recovery, Franka and legged go above harmonic extension. GC1–GC3 test this result against added baselines, imperfect observations and real data.

The revised PDF marks every change in blue, and all original numbers, tables, figures and equations keep their numbering. New items: Table R1 (baselines, §5.1), Figure R1 (Table 3 as curves, §5.3), Equations R1–R2 (§3.4), Table R2 (cost, Appendix A.8), Figure R3 (sampling seeds, Appendix C.8), Table R3 (DAVIS horse, Appendix D), Figure R2 and Table R4 (imperfect observations, Appendix H) and Table R5 (plausibility, Appendix I). The supplementary adds multi-view videos for all five families and a four-view video of the DAVIS horse. Following Reviewer 98K5, §4 and §5 now keep secondary numbers in the appendix and print Figure R1 at full width, with no change to any result.

GC1. Baselines (98K5, 5qQe, bPSV)

Table R1 update. The earlier Table R1 had two ARAP rows on different graphs, which made the comparison hard to read. All methods now run on the same graph for each family, and on the legged robots the graph's components are joined. ARAP has one row, converged to a relative change below 1e-5. The legged numbers therefore differ from the earlier version. CEPHEID has the lowest error in all five families.

Added baselines. We added thin-plate splines (TPS), biharmonic interpolation and iterative ARAP on all five columns, and ran DOMA on the two held-out families. Each cell is pooled error ÷ mean |F| (Eq. 6) on the conditions of Table 2. Lower is better. DOMA on DT4D is in Table 1, on its own split.

err ÷ mean |F|DT4DAMASSFrankacloth (held out)legged (held out)
harmonic0.26850.06930.10110.43770.2088
biharmonic0.18500.03920.07300.41790.2892
TPS0.21920.06110.04630.46180.2266
ARAP (converged)0.17570.02550.02630.39520.2435
DOMA———0.48950.8346
CEPHEID (ours)0.12430.02140.02010.36920.1786

CEPHEID has the lowest error on all five families. For cloth and legged, DOMA keeps the iteration budget and readout selected on DT4D. We did not tune it on the held-out families.

More recent cross-topology methods (bPSV). These methods address related problems but need inputs that sparse displacements do not provide:

methodrequired input that sparse displacements do not give
UniMate, AnyTop (UniMate also a text prompt)a skeleton
ScanMovea driving motion sequence
MeshLoomcomplete meshes at the observed keyframes (it registers them)
AnimateAnyMesha text prompt (generates motion rather than completing observed motion)
AuthorsOpenReview ↗

General response (2/2)

GC2. Imperfect observations (5qQe, bPSV)

CEPHEID stays below harmonic extension in most of the settings we test. All results use the frozen submitted model and its own amplitude recovery.

1. A real tracker on rendered video. We rendered four DT4D animal clips from their meshes (unlit, with procedural vertex colours), tracked them in 3D with TAPIP3D, using SC-GS depth as its depth input, over three depth seeds. Displacements run from the ground-truth first frame to frames 4–80, so they include accumulated tracker drift and first-frame error. The resulting observation error is large. Its RMS is 0.37–1.98× the mean motion |F|, or 0.25–1.24× after removing the smooth component (k = 12). It is also heavy-tailed: the per-vertex median ÷ RMS is 0.18–0.23 on moose, elk and raccoon and 0.49 on tiger, against 0.99 for i.i.d. Gaussian noise. We measure error on never-visible vertices, divided by the frame-0 bounding-box diagonal. CEPHEID is 28–30% below harmonic extension on moose (0.701, tracker error 0.37×), elk (0.716, 0.57×) and raccoon (0.700, 1.19×). It ties on tiger (1.007, seeds 0.984–1.025), where the tracker error is 1.98× and the amplitude features are clamped on 19 of 20 frames. On the DAVIS horse, CEPHEID agrees better than harmonic extension with 51 held-out CoTracker3 tracks in all 10 random input subsets, a ratio of 0.60 (two-level 95% CI [0.564, 0.635]).

2. Missing points. With up to 50% of observations dropped, CEPHEID stays below harmonic extension in all five columns. The largest ratio is 0.933, on legged.

3. Controlled corruptions. Both methods receive the same corrupted input. Under correlated heavy-tailed noise, CEPHEID stays below harmonic extension up to 0.71× |F| on DT4D, 0.40× on AMASS, 0.81× on Franka and 0.79× on legged. On cloth it is 2% worse at 0.84×. The table gives pooled error ÷ mean |F| (Eq. 6) under persistent spatially correlated offsets (size × |F|) and identity switches of ~10% of points. Both methods use the same 50 pairs per family on the graphs of Table 2, so the clean row differs slightly from Table R1.

DT4DAMASSFrankaclothlegged
cleanharmonic0.25070.07570.10570.42080.2458
CEPHEID0.10850.02270.02310.35900.2171
offset 0.1×harmonic0.27490.13850.15170.44320.2863
CEPHEID0.14700.10370.09830.39240.2693
offset 0.25×harmonic0.32490.25710.25590.50430.3703
CEPHEID0.21270.23830.22790.47270.3611
offset 0.5×harmonic0.42560.46530.44280.62890.5286
CEPHEID0.33540.46430.44580.63270.5262
identity switchharmonic0.33890.13870.16880.46330.3355
CEPHEID0.19540.07200.06250.42140.2781

For legged at 0.25× and AMASS, Franka, cloth and legged at 0.5×, the 95% CI of CEPHEID ÷ harmonic contains 1. Appendix H (Figure R2, Table R4) has the complete curves, including i.i.d. noise.

GC3. Real data and visualization (all reviewers)

DAVIS horse (Appendix D, Table R3). CoTracker3 gives 63 tracks. The paper's figure uses 12 as input, and we hold out the other 51. We project the completion into the video and measure the 2D distance to each held-out track wherever the tracker marks it visible. Mean error is 4.5 px for CEPHEID, 8.7 px for harmonic extension and 24.0 px for a rigid shift. CEPHEID is lower on 50 of 51 tracks and on all 35 frames. The tracker itself is not ground truth.

The paper's input subset turned out easier than random ones for both methods, so our main number uses random subsets. We drew 10 random 12-track input subsets from the 38 tracks occluded in at most 3 frames. CEPHEID averages 7.36 px against 12.30 px and is lower in all 10. The difference is −4.95 px (two-level 95% CI [−5.52, −4.50]), a ratio of 0.598. Drawing the subsets from all 63 tracks gives 5.88 against 9.93 px (−4.06, CI [−4.38, −3.72]), again lower in all 10. Each subset is scored on its own non-input tracks, so the means of the two pools should not be compared with each other.

Real 4D scans. Dynamic FAUST registrations are real 4D scans with SMPL topology like AMASS. We use the original D-FAUST registrations, not AMASS's refits. Our AMASS training split is the official one (nine subsets) and contains no D-FAUST data. The D-FAUST subjects also appear in MoSh, which is a validation subset and is never trained on. On D-FAUST, CEPHEID is 50.4% below harmonic extension (ratio 95% CI [0.480, 0.515]).

Multi-view videos (supplementary). Videos show ground truth, harmonic extension and CEPHEID from several views, and comparison/ adds ARAP side by side, each on one shared error colour scale. The README states how the conditions were chosen. A four-view video of the DAVIS horse has no ground truth, so it has no error colouring.

AuthorsOpenReview ↗

General response (3/3): presentation

We have revised the paper and our responses for readability.

  • Every new table (R1–R5) follows the layout of the main tables: one row per method, one column per family, absolute values, and a caption that states its main result.
  • Table R1 compares all methods on the same graph for each family, with one ARAP row.
  • §3.4 states the training objective and the sampler as Equations R1–R2.
  • Table 3 is now Figure R1, a full-width plot of the trend across densities.
  • §4 and §5 move secondary numbers into the appendix sections that already cover them.
  • The supplementary adds side-by-side videos of ground truth, harmonic extension, ARAP, and CEPHEID for every family.

The authors designed every experiment, checked every reported number against its result files, and take full responsibility for the method and the results.

Thank you for your time. To make the difference easier to see, we updated the supplementary videos: all methods are now drawn in grey, and a red box in each clip marks the unobserved region where the error gap is largest, chosen from per-vertex error.

  • DT4D (videos/00–03, comparison/dt4d_1–3): the legs and head. In videos/00, harmonic extension places the elk's head and legs away from the ground truth.
  • AMASS (videos/04–07, comparison/amass_1–3): the leg, foot or hand with no observed point.
  • Cloth (videos/12–14, comparison/cloth_1–3): the corners and folded edges far from the observed points.
  • Franka (videos/08–11, comparison/franka_1–3): the links between the observed points. The arm moves a few millimetres between frames, and at that scale CEPHEID's error is 80% lower than harmonic extension (Table R1).
  • Legged robots (videos/16–18, comparison/legged_1–3): the feet of the unobserved legs. On these rigid links, the case ARAP is built for, CEPHEID's error is 27% below converged ARAP and 15% below harmonic extension (Table R1).
  • Learning-based methods (learning_based/): NSDP and DOMA next to CEPHEID. DOMA optimizes a separate model for every scene, while CEPHEID is a single model applied without retraining. The difference is visible in shape: DOMA bends rigid parts and adds ripples to flat cloth, while CEPHEID keeps them as in the ground truth. learning_based/sequences/ shows one clip per family over consecutive frames: DOMA warps the robot body, bends the foot and inflates the cloth, NSDP crumples the surface, while CEPHEID stays close to the ground truth. CEPHEID is below DOMA on cloth (0.369 vs 0.490) and the legged robots (0.179 vs 0.835) (Table R1), and below both NSDP and DOMA at every density on DT4D (Table 1).

We hope the boxes make the difference visible.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.