acceptodds
Under review as a conference paper at ICLR 2027

OUTPUT-AWARE CONFORMAL GEOMETRY FOR STRUCTURED NONLINEAR PARAMETRIC MODELS

Abstract

Structured parametric models generate complete meshes from compact parameters, but conformal coverage alone does not control the spatial extent of a decoded prediction set. We introduce output-aware conformal geometry (OACG), which shapes parameter-space ellipsoids using the matrix geometric mean (GM) of residual precision and decoder sensitivity. Independent calibration and exact decoding give finite-sample marginal coverage of representable references under exchangeability, regardless of linearization accuracy. For full-rank outputs, the unregularized GM core minimizes a pooled-moment surrogate for linearized RMS extent up to a common scale. Rank-deficient joint outputs require a ridge; we compare one isotropic in parameter coordinates (GM-I) with one isotropic in residual-normalized coordinates (GM-R). At nominal coverage, both variants reduce decoded RMS extent against additive fusion in all five primary mesh settings. GM-R improves on additive fusion in all seven primary joint settings, whereas GM-I trails it on the three SMPL joint settings. From matched to participant-disjoint HOT3D evaluation, GM-I's extent increases by –, versus – for residual-only, additive, and GM-R geometries; population composition and participant overlap change together. A learned metric is smaller on EMDB but requires an additional geometry head. OACG makes fusion and ridge placement explicit choices for the spatial efficiency of structured prediction.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
conformal prediction, uncertainty quantification, structured outputs, human mesh recovery, 3D human pose and shape, SMPL, matrix geometric mean, decoder sensitivity

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_AzA4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Output-Aware Conformal Geometry (OACG) for constructing prediction sets for structured parametric models such as SMPL and MANO. The main idea is to combine prediction residual statistics with decoder sensitivity using a matrix geometric mean, so that the resulting parameter-space ellipsoids account for the spatial displacement of decoded outputs. The authors also consider two regularization strategies, GM-I and GM-R, for rank-deficient outputs.

The paper provides a theoretical analysis of the proposed geometry and evaluates it on 3DPW, EMDB, and HOT3D with several body and hand reconstruction models. The results show reduced decoded RMS extent compared with additive fusion in most settings, although the relative performance of GM-I and GM-R varies across output types.

Critical strengths
  1. The motivation is reasonable. Parameter-space uncertainty does not necessarily reflect uncertainty in the decoded mesh. Using decoder sensitivity to construct the prediction region is a sensible way to address this issue.

  2. The paper includes a fairly thorough evaluation. I appreciate the comparisons across different predictors and output types, particularly the analysis of how regularization affects mesh and joint predictions. The authors are also careful to distinguish the theoretical surrogate result from the actual nonlinear decoded extent.

Critical weaknesses
  1. The theoretical result is somewhat limited. The optimality result applies to a linearized surrogate rather than the actual calibrated nonlinear extent. It is therefore unclear how much this result explains the improvements observed in practice. I would like to see a more direct connection between the theoretical objective and the reported evaluation metric.

  2. The improvements are not consistent across settings. While the results on full meshes are encouraging, GM-I performs worse than additive fusion on all three SMPL joint settings. GM-R performs better on joints, but the preferred variant still depends on the task. This makes it difficult to identify a general advantage of the proposed construction beyond selecting the best variant for each setting.

  3. The computational cost is relatively high. The method requires Jacobian computation and matrix square roots. Section 5.6 reports considerable overhead in some settings, especially for HOT3D meshes at larger batch sizes. Given that the improvements are sometimes modest, I am not fully convinced that the additional complexity is justified in all cases.

Questions for authors
  1. Can the authors provide an empirical comparison between the surrogate objective and the actual calibrated nonlinear extent? Does a lower surrogate value generally lead to a smaller decoded prediction set?

  2. Is there a practical way to decide between GM-I and GM-R without evaluating both on each new task? The results suggest that this choice can substantially affect performance.

  3. How does the runtime scale with the dimensionality of the parameter space? A direct end-to-end runtime comparison with the additive and learned-metric baselines would help clarify the practical trade-off.

3 comments
AuthorsOpenReview ↗

Response to Reviewer AzA4

Thank you for the questions on theory, variant selection, and cost. Extent comparisons use nominal 90% coverage. Page references are to the manuscript with the rebuttal appendix (Appendix C, pp. 44–64).

1. Reply to W1 & Q1: Connecting the surrogate to calibrated nonlinear extent

The surrogate and decoded extent give closely aligned rankings on 3DPW meshes. For joints, the rankings diverge after nonlinear decoding. Using the same residual statistics, we compare 110 candidates: 10 GM-I, 10 GM-R, and 90 additive configurations. We average each candidate’s results across splits, then measure how closely the surrogate and extent rankings agree using Spearman correlation (Appendix C.3.1, p. 46).

Surrogate versus exact squared extent3DPW mesh3DPW jointsEMDB joints
All 110 candidates0.97940.58280.5140
GM-I0.8788−0.4424−0.5030
GM-R0.5879−0.6485−0.6485

These are the ρ(U1,U4)\rho(U_1,U_4) entries of Table 30 (p. 47). The four-stage comparison in Eq. (42) and Table 29 (p. 46) identifies where the ordering changes:

Adjacent stages3DPW mesh3DPW jointsEMDB joints
Pooled surrogate → residual-moment surrogate0.99690.99460.9880
Residual-moment surrogate → calibrated linear extent0.99170.99470.9921
Calibrated linear extent → exact squared extent0.98020.57700.5232

Section 3.3 (p. 4) already identifies the joint-specific mechanism: 14 joints provide 42 coordinates for 82 parameters, leaving directions with zero first-order cost. Ridge controls perturbations in those directions, whose finite response can be nonlinear (Section 4.3, pp. 5–6). We measure that contribution on identical perturbations: five splits per joint task and 64 shared draws per input. Writing L=JδL=J\delta and r=g(δ)−g(0)−Lr=g(\delta)-g(0)-L, Eq. (44) (p. 46) gives

E^2=E^∥L∥W2+2E^⟨L,r⟩W+E^∥r∥W2.\widehat E^2=\widehat{\mathbb E}\|L\|_W^2+2\widehat{\mathbb E}\langle L,r\rangle_W+\widehat{\mathbb E}\|r\|_W^2.

For GM-R on 3DPW joints, Table 31 (p. 48) yields:

Ridge λ\lambdaAll-row linear extent (mm)Exact extent (mm)Linear shareCross shareRemainder share
10−510^{-5}67.71,370.70.24%−0.49%100.24%
0.0167.8300.85.08%−4.95%99.87%
0.5, selected on development data72.577.088.37%−11.02%22.64%

Selecting ridge by exact decoded extent gives a consistent choice and improves joint extent. All ten 3DPW joint deletion folds select λ=0.5\lambda=0.5 (Table 41, p. 59). The full development-selected GM-R configuration reduces joint extent from additive’s 83.18 to 78.13 mm (Table 42, p. 60).

The small-ridge results explain why this selection matters. The all-row linear estimate stays near 68–73 mm, while nonlinear terms drive exact extent to 1,371 mm; shape-only perturbations already produce 1.29 m (Table 31, p. 48). Exact extent closely tracks the nonlinear remainder (ρ=0.90\rho=0.90–0.990.99; Table 33, p. 49). This decomposition locates the ranking reversals in the finite-radius nonlinear response and quantifies the correction in Proposition A.4 (p. 20) and Eq. (45) (p. 47).

AuthorsOpenReview ↗

Response to Reviewer AzA4 Continued

2. Reply to W2 & Q2: A general advantage and a practical variant-selection rule

GM-R gives positive mean paired gains over additive in all 12 settings, with a median gain of 11.09%. Eleven meet the prespecified criterion of ≥5% mean gain and ≥16/20 positive splits (Table 28, p. 44). Variant selection remains useful: GM-I trails additive on the three SMPL joint settings but outperforms GM-R on the original participant-disjoint HOT3D mesh and RjointsR_{\mathrm{joints}} outputs (Table 2, p. 8).

Our restricted search, R490 + I10, retains the full-search choice at about half the development cost. It searches 490 GM-R configurations, then tests ten GM-I ridge strengths using the selected residual-moment hyperparameters. It chooses the smaller exact development extent, refits and freezes the configuration, and calibrates independently (Section 3.4, pp. 4–5; Appendix C.5.3, p. 59).

This reduces the search from 980 to 500 candidates and reproduces the full search’s final variant and configuration in 28/28 runs (Table 47, p. 62). Actual 3DPW executions confirm the same choices and reduce total development time by 48.8% for both mesh and joints (Tables 48–49, p. 62).

3. Reply to W3 & Q3: Total development cost and parameter dimension

OACG constructs geometry analytically, without training an additional geometry network (Eqs. (6)–(7), p. 4). We compare total development time from cached inputs to the final fitted configuration, including search, exact-extent evaluation, and learned C’s training and checkpoint selection. All methods use the same A100, eight CPU threads, and 2,000 cached 3DPW inputs (Appendix C.5.4, p. 61).

Total time (minutes)Full OACGRestricted OACGLearned C (estimated)
Mesh40.3820.66110.22
Joints40.0420.49110.29

OACG totals are measured. C’s total is estimated by scaling measured training and checkpoint-evaluation costs to 80 configurations × five rotations, then adding measured final training (Tables 49–50, pp. 62–63).

Including training, full OACG reduces total development time by 63.4%–63.7%; restricted OACG increases the savings to 81.3%–81.4% while preserving the selected configurations. At nominal 90%, selected GM-R also achieves smaller extent than rebuilt C: 94.46 vs. 99.21 mm for mesh and 78.13 vs. 78.62 mm for joints (Table 42, p. 60; Appendix C.2, p. 45).

For dimension scaling, CPU GM matrix construction at batch size 1 takes 0.432, 2.495, and 20.240 ms/input at d=16,82,256d=16,82,256, respectively. Tables 53–54 (p. 64) report the full matrix-cost scan for batch sizes 1 and 128; the development totals above use SMPL’s actual dimension, d=82d=82.

AuthorsOpenReview ↗

Summary of updates addressing W1–W3 and Q1–Q3

Thank you again for your questions on the theoretical explanation, variant selection, and computational cost. We have added targeted experiments and posted detailed responses. The main findings are: W1 / Q1 — Surrogate versus nonlinear extent. We compare 110 candidates across 20 calibration splits. On 3DPW meshes, surrogate and exact-extent rankings agree closely (Spearman ρ = 0.9794). For joints, the four-stage comparison locates the ranking reversals at nonlinear decoding. A matched-perturbation decomposition shows that small-ridge extent growth comes from nonlinear terms; at the development-selected ridge, these terms are controlled and exact extent remains close to the linear scale (Tables 29–33, pp. 46–49). This explains the roles of the analytical GM core and exact-extent-based ridge selection.

W2 / Q2 — General gains and practical selection. GM-R achieves positive mean paired gains over additive in all 12 settings, with a median of 11.09% (Table 28, p. 44). Our R490 + I10 procedure searches 490 GM-R configurations and only ten GM-I ridge strengths. It reduces the candidate budget from 980 to 500, consistently preserves the full search's optimal choice in 100% of replayed runs (28/28), and reduces measured total development time by approximately 48.8% on both 3DPW outputs (Tables 47–49, p. 62).

W3 / Q3 — Training-inclusive cost and dimension scaling. OACG is fully plug-and-play, requiring no additional geometry-network training. Restricted OACG reduces total development time by 81.3%–81.4% relative to learned C’s estimated training-inclusive cost (Tables 49–50, pp. 62–63). We also added the requested unified full-pipeline comparison and a controlled matrix-cost scan over d = 16 ∼ 256 at batch sizes 1 and 128 (Tables 51, 53–54, pp. 63–64).

We would particularly welcome your assessment of the surrogate analysis and the practical selection procedure. Please let us know whether any concerns remain; we would be happy to discuss them further.

Reviewer_fb2pOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes Output-Aware Conformal Geometry (OACG), a framework for constructing calibrated prediction regions for structured nonlinear parametric models (SMPL and MANO). The authors identify that conventional conformal prediction guarantees marginal coverage but does not control the spatial extent of the resulting decoded meshes or joints. To address this limitation, OACG constructs parameter-space ellipsoids by combining residual precision and decoder sensitivity through their matrix geometric mean, avoiding the explicit mixing coefficient required by additive fusion. The residual precision is estimated from prediction errors, while decoder sensitivity is obtained from the Jacobian of the nonlinear output mapping. Since rank-deficient outputs can produce singular sensitivity matrices, the authors investigate two regularization strategies: GM-I, which applies an isotropic ridge in parameter coordinates, and GM-R, which applies an isotropic ridge in residual-normalized coordinates. The geometry and its hyperparameters are selected using development data, while independent conformal calibration determines the ellipsoid radius. Exact nonlinear decoding then defines the output prediction set, providing finite-sample marginal coverage of representable references under exchangeability. The authors also theoretically establish that the unregularized geometric-mean core minimizes a pooled linearized RMS-extent surrogate under full-rank conditions. Quantitative experiments are conducted using HMR2.0, ProHMR, and HaMeR with SMPL and MANO on 3DPW, EMDB, and HOT3D. The evaluation includes comparisons with isotropic, residual-only, sensitivity-only, additive-fusion, and learned-metric baselines. At nominal 90% coverage, both GM variants reduce decoded RMS extent relative to additive fusion across all five mesh settings, while GM-R achieves improvements across all seven joint settings. Additional analyses examine the effects of ridge placement and strength, nonlinear decoding, computational cost, and population differences between matched and participant-disjoint HOT3D settings. The results indicate that GM-I exhibits smaller relative changes in decoded extent across the observed HOT3D population designs, whereas GM-R provides more consistent improvements for joint outputs.

Critical strengths

1. Well-motivated problem formulation.

The paper identifies an important limitation of conventional conformal prediction for structured nonlinear parametric models: guaranteeing coverage does not necessarily produce spatially compact prediction sets. Optimizing decoded RMS extent rather than parameter-space volume is particularly meaningful for 3D meshes and articulated structures.

2. Solid theoretical analysis.

The authors establish finite-sample marginal coverage for complete decoded prediction sets under exchangeability and representability assumptions. They also derive the optimality of the geometric-mean core for a pooled linearized RMS-extent surrogate under full-rank conditions, while explicitly distinguishing this result from optimality under exact nonlinear decoding. The proposed matrix geometric mean combines residual precision and decoder sensitivity without requiring an explicit mixing coefficient, unlike additive fusion. The formulation provides a mathematically motivated approach to constructing output-aware prediction regions.

Critical weaknesses

1. Insufficient qualitative evaluation and spatial error visualization.

Despite the extensive quantitative evaluation, the qualitative results are limited to a small number of examples in Figures 3 and 4 of the appendix, with only two body and two hand examples. These visualizations compare residual-only, additive fusion, and GM-I, but omit GM-R, making it difficult to visually assess the differences between the two proposed variants. Although Figure 3 includes per-vertex RMS displacement heatmaps, these measure displacement from the point prediction rather than reconstruction error relative to the ground truth. The authors should provide more comprehensive qualitative comparisons across different datasets, including challenging occlusion cases and failure examples. Per-vertex error maps relative to the ground-truth mesh, together with spatial uncertainty or displacement maps, would help reveal where the proposed geometry improves or degrades reconstruction accuracy and prediction-set compactness. Side-by-side visualizations of additive fusion, GM-I, and GM-R would make the reported quantitative improvements more interpretable.

2. Inconsistent performance across output representations.

Although both GM variants improve mesh prediction, their relative performance is highly task-dependent. GM-I performs worse than additive fusion across all three SMPL joint settings, with paired losses ranging from 6.4% to 34.7%. GM-R generally performs better on joints but is substantially worse than GM-I on participant-disjoint HOT3D meshes. These results suggest that the proposed geometry requires task-specific selection. The learned-metric baseline outperforms both GM variants on EMDB mesh and joint prediction, with particularly notable improvements on joints.

3. Computational overhead and incomplete end-to-end wall-clock evaluation.

OACG requires computing a decoder Jacobian and matrix square roots for each input, introducing additional computational overhead compared to additive fusion. According to Appendix B.6, the geometry-construction overhead reaches 13.7% for 3DPW meshes, 36.0% for 3DPW joints, and 500.7% for HOT3D meshes at batch size 128. The development-stage hyperparameter selection also incurs substantial computational costs, requiring 11.6 GPU-hours for GM-I and GM-R combined on 3DPW meshes and 2.5 GPU-hours per HOT3D task. Although these costs are lower than those of additive fusion due to its larger hyperparameter search space, the additional per-input computation remains a practical concern. More importantly, the reported wall-clock measurements exclude predictor inference and sampled mesh decoding, and the learned-metric baseline is not evaluated under the same runtime protocol. The paper lacks a comprehensive end-to-end wall-clock comparison, making it difficult to assess whether the improvements in decoded RMS extent justify the additional computational cost in practical applications.

Questions for authors
  1. The paper cites only 22 references, which seems rather limited given the extensive literature on conformal prediction, uncertainty estimation, and geometry-aware representations.
6 comments
AuthorsOpenReview ↗

Response to Reviewer fb2p Part 1

Thank you for your detailed comments. We have expanded our spatial evaluation, analyzed new-participant selection, and added a training-inclusive cost comparison in the revised manuscript (new rebuttal studies in Appendix C, pp. 44–64).

1. Reply to W1: Spatial visualization, ground-truth error, and challenging cases

Figures 11 (3DPW, p. 55) and 12 (EMDB, p. 57) compare ground-truth error and uncertainty maps for additive, GM-I, GM-R, and learned C. They show spatial improvements and failure cases, including an atypical 3DPW example where GM-R has larger decoded extent than additive.

Across 20 calibration splits, GM-R reduces aggregated RMS displacement at every output point across all four HMR2.0 settings, with the feet accounting for ~35% (3DPW) and ~33% (EMDB) of the whole-mesh squared-extent reduction (Appendix C.4.2, p. 51; Figure 9 and Table 35, p. 53).

Because smaller regions must be evaluated alongside reference inclusion, we pair extent with reference-parameter certification (S≤q^S\leq\widehat q, which guarantees decoded reference inclusion via Theorem 4.1, p. 5) on the same input groups at nominal 90% coverage:

Input groupExtent typeGM-R / C extent (mm)GM-R / C certification
3DPW mesh, pose P90–98Linearized89.4 / 111.582.0% / 75.9%
3DPW joints, pose P90–98Linearized74.9 / 85.483.9% / 78.5%
EMDB mesh, pose ≥P98Exact110.0 / 115.159.1% / 39.2%
EMDB joints, pose ≥P98Exact96.5 / 91.064.0% / 40.2%

For EMDB, we exactly decode all atypical-group evaluation rows across 20 splits (64 common draws/row) against an equally sized typical control using matched rows and weights (Appendices C.4.3–C.4.4, pp. 51, 54; Tables 36, 38, pp. 55–56). Conversely, in the highest-error 10% of 3DPW mesh inputs, GM-R / CC yield linearized extents of 88.0 / 104.5 mm and certification rates of 25.4% / 36.4% (Table 36, p. 55), cleanly separating pose atypicality from reconstruction difficulty.

AuthorsOpenReview ↗

Response to Reviewer fb2p Part 2

2. Reply to W2: Performance across outputs and learned C

GM-R improves on additive in all 12 primary settings without training an additional geometry network. At nominal 90%, its paired gains have median 11.09%; 11 settings meet the prespecified criterion of ≥5% mean gain and ≥16/20 positive splits (Table 28, p. 44). GM-I trails additive on the three SMPL joint settings, where the pullback has rank ≤42 in d=82d=82 and the ridge supplies the entire kernel penalty (Section 5.3, p. 8).

On HOT3D-Clips, GM-R's deficit to GM-I on new-participant mesh disappears under participant-grouped development. In the participant-disjoint HOT3D design, GM-R/GM-I mesh extents are 38.45/26.68 mm (Table 2, p. 8). HOT3D-Clips has 24 development sequences from three participants and 99 deployment sequences from six others. Sequence-based selection reproduces the gap; participant-grouped selection removes it:

Mesh, development protocolAdditiveGM-IGM-R
Sequence-based39.7127.1638.39
Participant-grouped31.4226.3126.39

Extents are in mm at nominal 90%; additive is reselected under each protocol (Table 46, p. 61). Grouping changes GM-R's selected outer ridge from 0.5 to 0.001 (Table 45, p. 61). Grouped development selects GM-I for mesh, with a 16.35% paired gain over additive (20/20 positive splits). For RjointsR_{\mathrm{joints}} and RdistalR_{\mathrm{distal}}, both protocols select GM-R, with grouped gains of 14.07% and 12.22% (20/20 each; Table 44, p. 60). We therefore recommend participant-grouped development when deployment involves new participants (Section 6, p. 10). The original disjoint pool was not rerun under grouping (Appendix C.5.2, pp. 58–59).

C conditions its metric on image features; OACG uses pooled residual statistics and decoder geometry (Section 6, p. 10). C has smaller overall EMDB extent. The archived results below are in mm at nominal 90% (Table 1, p. 7):

OutputAdditiveGM-IGM-RC
Mesh123.44109.03107.5799.02
Joints103.54109.6494.5281.28

Exact decoding by pose group tests where this comparison changes. Below are GM-R's mean paired percentage differences from additive (add.) and C; negative means smaller GM-R extent. Parentheses count splits with smaller GM-R extent.

Pose groupMesh vs add.Mesh vs CJoints vs add.Joints vs C
Typical control−13.0% (20/20)+11.8% (1/20)−9.1% (19/20)+18.0% (0/20)
P90–98−10.2% (20/20)−0.3% (11/20)−6.0% (16/20)+9.9% (1/20)
≥P98−7.5% (16/20)−3.3% (12/20)−4.4% (14/20)+6.7% (7/20)

GM-R has smaller mean extent than additive in every pose group. Relative to C, mesh extent changes from 11.8% larger on typical inputs to 3.3% smaller in the most atypical group. The joint gap narrows from 18.0% to 6.7%, with C still smaller (Figure 10, p. 55; Table 38, p. 56).

AuthorsOpenReview ↗

Response to Reviewer fb2p Part 3

3. Reply to W3: Computational cost, including training

OACG constructs geometry analytically without training an additional geometry network. We compare total development time from cached inputs to the final fitted geometry. This includes candidate construction or training, temporary calibration, exact-decoder evaluation, selection, and final fitting.

Table 49 (p. 62) compares the methods on the same A100, eight CPU threads, and 2,000 development inputs. Times are in minutes:

3DPW outputFull OACGRestricted OACGLearned C (estimated)
Mesh40.3820.66110.22
Joints40.0420.49110.29

Full OACG searches 980 configurations; the restricted procedure searches 490 GM-R configurations plus ten GM-I ridge strengths and retains the same final configurations. OACG totals sum measured search and final-fit times. (Appendix C.5.4, p. 61; Table 48, p. 62, and Table 50, p. 63).

The full OACG procedure reduces total development cost by 63.4–63.7% relative to C's estimate; restricted selection increases this reduction to 81.3% for mesh and 81.4% for joints. The retained GM-R configurations reduce mean exact extent from 105.67 to 94.46 mm on mesh and from 83.18 to 78.13 mm on joints, corresponding to 10.61% and 6.07% reductions in the means (Table 42, p. 60, “All ten” rows). The same configurations deliver both the smaller development budget and the spatial-efficiency gains.

Reply to Q1: Coverage of related work

Thank you for this suggestion. Since the limitation of pages, we just cited the most related works at first. We have expanded the bibliography from 22 to 35 references and strengthened Section 2 (pp. 2–3). The additions address the three areas you highlighted:

1. Conformal prediction

  • Johnstone and Cox (2021), Conformal Uncertainty Sets for Robust Optimization.
  • Tumu et al. (2024), Multi-modal Conformal Prediction Regions by Optimizing Convex Shape Templates.
  • Fang et al. (2025), CONTRA: Conformal Prediction Region via Normalizing Flow Transformation.

These works broaden the discussion of multivariate conformal regions, including ellipsoidal constructions, optimized convex templates, and regions obtained through learned transformations. The revised comparisons clarify how OACG’s decoded-displacement objective relates to these approaches to region construction and efficiency.

2. Uncertainty estimation

  • Sankaranarayanan et al. (2022), Semantic Uncertainty Intervals for Disentangled Latent Spaces.
  • Teng et al. (2023), Predictive Inference with Feature Conformal Prediction.
  • Tang et al. (2025), Accelerating Feature Conformal Prediction via Taylor Approximation.

These additions address uncertainty estimation in latent and feature spaces, including semantic-coordinate intervals, calibrated feature regions, and gradient-based approximations. We explain their connections to OACG’s construction of parameter regions and propagation of uncertainty through a nonlinear decoder.

3. Geometry-aware representations

  • Arvanitidis et al. (2018), Latent Space Oddity: On the Curvature of Deep Generative Models — decoder-induced geometry and distortion of latent-space distances.
  • Yang and Pavone (2023), Object Pose Estimation With Statistical Guarantees: Conformal Keypoint Detection and Geometric Uncertainty Propagation — propagation of calibrated keypoint uncertainty into geometric pose sets.
  • Kubo and Ando (1980), Means of Positive Linear Operators — the mathematical foundation for the matrix geometric mean used in our metric construction.

These references connect decoder geometry, geometric uncertainty propagation, and matrix fusion to OACG’s combination of residual statistics and decoder sensitivity.

Could you suggest any additional closely related papers in these areas that we have not yet cited? We would welcome your recommendations to further improve the discussion.

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

This paper proposes Output-Aware Conformal Geometry (OACG) for uncertainty estimation with fixed structured predictors. It combines historical parameter errors and decoder sensitivity through a matrix geometric mean to construct parameter-space ellipsoids. Independent conformal calibration determines their radius, and exact decoding maps them to sets of complete meshes or joint configurations.

Critical strengths
  • The paper addresses an important distinction between coverage and spatial efficiency.
  • Combining residual information with decoder sensitivity is a reasonable approach. The method applies to fixed predictors without training an additional geometry network.
  • The paper clearly separates the conformal coverage guarantee, the optimality of a linearized surrogate, and the empirical behavior of nonlinear decoded sets.
Critical weaknesses
  • The main metric measures average displacement under uniform sampling in the parameter ellipsoid. It does not control extreme decoded outputs or establish tighter spatial envelopes.
  • On 3DPW, only ten development sequences are used to select among hundreds of configurations.
Questions for authors

Please see the weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer gzte

We sincerely appreciate the reviewer’s follow-up comments. We address the raised concerns below. The new studies are collected in Appendix C (pp. 44–64), and we will integrate them into the main presentation after the rebuttal period.

1. Reply to W1: Average displacement, sampled tails, and extreme outputs

The paper's RMS extent is based on average squared decoded displacement under uniform parameter-region sampling (Eq. (3), p. 4). To establish which spatial benefits this captures, we add tail quantiles and extreme searches that directly probe large deformations.

The new measurements show that GM-R's gains on 3DPW meshes extend from average displacement to sampled tails: relative to additive fusion, RMS decoded extent decreases by 9.6%, the 99th percentile of RMS displacement decreases by 5.5%, and the 99th percentile of maximum-vertex displacement decreases by 13.3%. Appendix C.4.1 (p. 51) specifies the protocol, and Table 34 (p. 52) reports the complete comparison.

At nominal 90% coverage, we use five calibration splits, 24 evaluation frames per sequence, and 1,024 shared unit-ball draws per input. Each method transforms these draws using its own geometry and calibrated radius. Tail quantiles are computed per input and averaged with equal sequence weights. The following values are in mm:

Output and metricAdditiveGM-IGM-R
Mesh: RMS decoded extent101.793.992.0
Mesh: RMS displacement, q99138.4135.0130.8
Mesh: maximum-vertex displacement, q99417.7370.5362.1
Joints: RMS decoded extent80.699.676.9
Joints: RMS displacement, q99107.9156.6109.0
Joints: maximum-joint displacement, q99249.6347.0260.9

For joints, GM-R reduces average extent by 4.5%; the two q99 values increase by 1.0% and 4.6%. Thus the mesh improvement reaches the sampled tails, and the joint improvement is concentrated in average displacement.

We search for extreme displacement with projected Adam. The searched RMS values for additive/GM-R are 255.7/266.1 mm on meshes and 190.2/252.1 mm on joints. These attained values are lower bounds on each region's maximum. The evidence locates the mesh gains in average and sampled-tail behavior; GM-R has larger attained search values.

2. Reply to W2: Development allocation and selection stability

Ten development sequences remain after reserving disjoint data for final calibration and evaluation. Per split, our 60-sequence 3DPW protocol allocates 40 sequences to calibration and 10 to evaluation, leaving 10 fixed to determine the geometry and hyperparameters before final radius calibration (Appendix B.3, p. 23).

Leave-one-sequence-out re-selection selects GM-R and preserves positive mean GM-R gains across all ten folds for both outputs (Appendix C.5.1; Table 39, p. 58). Each fold reruns the five-rotation search (4,410 additive and 490 per-GM-family configurations) on nine sequences, refits winners on all nine, and evaluates 90%-coverage extent reductions against the same-fold additive baseline across 20 common splits. Our prespecified consistency criterion requires a ≥5%\ge 5\% mean paired gain and positive gains on ≥16/20\ge 16/20 splits:

OutputGM-I paired-gain rangeGM-R paired-gain rangeGM-R consistent folds
Mesh+6.21% to +9.13%+9.03% to +11.35%10/10
Joints−23.46% to −20.65%+5.09% to +7.12%9/10

All mesh folds select the identical GM-R configuration, while joint folds consistently retain GM-R (λ=0.5\lambda=0.5) with varying residual-moment hyperparameters (Table 41, p. 59). The sole joint exception (omitting capoeira) still achieves a +5.30% mean gain across 15/20 positive splits. At 95% coverage, 9/10 mesh and 5/10 joint folds meet this criterion (see Tables 40–43, pp. 58–60 for full per-fold details).

AuthorsOpenReview ↗

Response to Reviewer gzte — Continued

3. Supporting W1–W2: Methodological foundations and verification

OACG builds on conformal calibration, constructing its geometry from residual statistics and decoder Jacobians without training an additional geometry network. Its theoretical basis and empirical evidence are directly verifiable:

  • Coverage and geometric efficiency. Following Messoudi, Destercke, and Rousseau (2022, Sections 2.3 and 3.2), who decouple coverage validity from region-size efficiency (ellipsoid volume), OACG freezes its geometry before independent final calibration. Under conditional exchangeability and reference-reconstructibility (Theorem 4.1, p. 5; proof in Appendix A.1, p. 15), exact decoding transfers parameter-ellipsoid coverage directly to the decoded output region. While Proposition A.2 (p. 16) shows that worst-case displacement and volume proxies can favor different geometries in a fixed linearized setting, our average, tail, and extreme-search metrics provide the empirical counterpart after nonlinear decoding.
  • End-to-end selection evaluation. Following Cawley and Talbot (2010, Sections 5–6)—who treat model selection as an integral part of the evaluated fitting procedure—our leave-one-sequence-out analysis repeats geometry estimation, full-grid search, refitting, and independent calibration from scratch per fold (full results in Tables 40–43, pp. 58–60).
  • Independent sampling check. Repeating tail estimation with an independent 1,024-draw stream on the first 24 frames per split (Table 34, p. 52) yields median inter-stream relative differences of 1.91% for RMS q99 and 2.52% for maximum-vertex q99 on GM-R meshes, confirming sampling repeatability.

Thank you again for your effort and review.

My concerns have been addressed. Therefore, I keep my score.

Thank you for reviewing our responses and confirming that your concerns have been addressed. We appreciate your time and constructive input.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.