acceptodds
Under review as a conference paper at ICLR 2027

Align Before You Correct: Surface-Aligned Evidence for 3D Face Refinement

Abstract

Refining an existing 3D face reconstruction requires choosing where corrective evidence should be represented. Camera-ray corrections vary with viewpoint, while coefficient updates depend on a specific face model. We therefore express the correction on the reconstructed surface. We project each surface point into the calibrated views, where visible appearance and viewing geometry are fused to predict a signed displacement along the local normal. A robust parametric fit converts these displacements into face-model coefficients. Surface displacements are easier to predict than both ray-space targets. On held-out FaceScape identities, the advantage is smallest for frontal views and largest for oblique ones. Residual correlation is 0.39 for ray projection and 0.48 for ray intersection, compared with 0.52 on the surface using one view and 0.74 using five. With five views, symmetric Chamfer error drops by 4.5% from a strong start and by 8.5% from an image-predicted start. This recovers half of the reduction achieved by ground-truth displacements through the same fit and roughly twice the gain from averaging one-view predictions. Direct coefficient regression from the same evidence produces about one fifth of the gain from the image-predicted start, while population-mean, constant, and random displacement fields provide no comparable improvement. Without retraining, the same predictor reduces error by 9.5–11.7% for four published reconstructors and a multi-view landmark fit on FaceScape and by 4.3% for Pixel3DMM on Headspace. Expressing correction on the reconstructed surface therefore allows the same image evidence to transfer across calibrated rigs, face models, and starting reconstructors.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D face reconstruction, multi-view representation learning, 3D computer vision, geometric correction, 3D morphable models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8c4FOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a surface aligned residual representation for refining existing 3D face reconstruction from multiple calibration images. This method does not predict the correction along the camera light or directly update the facial model coefficients, but represents the correction of each sampled surface point as a signed displacement along its local normal. The image features extracted by the frozen DINOv2 encoder are projected onto the currently reconstructed surface, combined with the observed geometric shape, and aggregated between views using mean and standard deviation pools. Learn the decoder to predict the residual displacement field, and then convert it into parameterized face model coefficients through robust regularization fitting.The core contribution is a model independent calibration interface that separates residual prediction based on learning images from downstream parameterized geometric representations. Experiments on FaceScape show that surface normal correction is superior to camera ray correction of targets, especially in oblique viewing angles. Through five views, this method achieved a residual correlation of 0.738 and reduced the symmetric chamfer distance by approximately 4.5% from strong initialization and 8.5% from image prediction initialization.The author also evaluated several existing face reconstruction methods and demonstrated the transfer of cross dataset to Headspace without retraining the residual predictor.

Critical strengths

1.Clear and innovative principle expression, sufficient geometric foundation Anchoring corrective evidence onto surface normals is a simple but important contribution, rather than a simple engineering improvement. By transferring the correction space from the perspective dependent ray space and the coefficient space bound to the model to the surface relative space, this method achieves dual decoupling: parametric decoupling from camera perspective and facial models. The geometric analysis in Section 3.1 strictly explains the reason why the normal displacement condition is better, and the theoretical intuition is highly consistent with the experimental results. This insight from the representational level can be extended beyond facial refinement, and has reference significance for broader multi perspective 3D refinement research. 2.Strict experimental design, complete control and ablation experiments The experimental design of this article is rigorous and the control of confounding variables is very effective. The author controlled the initial reconstruction quality by interpolating the initial values, isolated the influence of the features themselves through parameter matching, and conducted comprehensive ablation experiments on variables such as encoder, fusion strategy, and model parameterization. Of particular importance, the paper established strong negative controls (overall mean field, random field, other subject images, etc.), ruling out other possible explanations. The zero sample transfer experiment across datasets also convincingly validated the reusability of the predictor.

Critical weaknesses
  1. Insufficient verification in real unrestricted scenarios This method is based on two strong premises: the camera has been accurately calibrated and the initial grid is well aligned. Although the appendix tested the effects of camera perturbations, it only covers independent single factor perturbations and assumes that the initial geometric shape is accurate. This paper did not evaluate performance in common real-world scenarios, such as uncalibrated cameras, large initial pose errors, or low-quality initial reconstructions with topological defects. Due to the dependence of the entire surface alignment framework on precise 2D-3D correspondence, the performance degradation under imperfect initial conditions is still unclear, which limits the applicability of this method in uncontrolled acquisition scenarios. 2.In cross dataset migration, the refinement effect of individual specificity is relatively weak On the headspace dataset, the vast majority of error reduction comes from correcting the overall level sharing bias of Pixel3DMM, while the additional gain brought by specific subject image evidence is very limited (about 0.03mm). Although the authors mentioned this phenomenon, they did not analyze its fundamental causes or demonstrate meaningful fine-grained individual geometric refinement effects across datasets. Therefore, the conclusion of "cross dataset generalization" is more suitable for correcting system reconstruction bias, rather than the more valuable and challenging refinement of individual geometric details in the scene, which weakens the practical impact of transferable conclusions.
Questions for authors

Q1. How does the proposed correction representation compare to the powerful matching grid alignment refinement alternative? Can the author evaluate a multi view mesh alignment refinement tool that predicts 3D or normal displacement of vertex directions based on the same projection features under matched training data, regularization, and optimization budget? This will help determine whether the proposed interface provides significant advantages beyond using established surface alignment features and geometric fitting.These pieces of evidence will greatly affect my evaluation of the novelty of this method. Q2. How sensitive are the results of external reconstructors to reference scan assisted alignment? Appendix F.2 states that prior to refinement, each external reconstruction is aligned with the FLAME fit exported from the scan. Can the author evaluate this method using initial alignment based solely on available images and reconstructed output estimates, without reference to scanning geometry? Or, can the author quantify the impact of initial placement errors? This will clarify to what extent the reported improvements have been translated into a realistic reconstruction pipeline. Q3. To what extent are the improvements on Headspace attributable to subject-specific image evidence rather than population-level correction? Considering that the overall average residual field performs almost as well as the predicted values of specific subjects, can the author analyze the improvement of specific subjects after considering this shared correction? Segmentation based on initialization errors, visible facial coverage, or anatomical regions will be particularly useful. The answer will clarify whether the benefits of cross datasets mainly represent system bias correction or personalized geometric refinement.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8c4F (1/2)

Thank you very much for the positive and careful review. We are glad that you found the geometric principle clear and the controls rigorous. Your three questions were specific and helpful, and we ran a new experiment for each of them. Section, table and figure numbers below refer to the revised PDF, where the changes for your points are marked in green and purple. Gains are Chamfer reductions on FaceScape in canonical units, where 1 c.u. is about 0.82 mm.

Q1. How does the representation compare with a matched vertex-displacement refiner?

Thank you for suggesting this experiment. We trained exactly this alternative, and Appendix B.1 reports it. The refiner uses the same projected DINOv2 features, the same fusion, the same training data and the same fit as our five-view model. Only the head differs. It predicts a 3D vertex offset, and the parameter count changes by 0.05%. The result is that the two heads perform almost the same, while targets outside the surface-aligned space are clearly weaker.

Target, with the same features and fitGain at α = 0.5Identities improved
Scalar normal offset, ours+0.090239 of 40
3D vertex offset+0.095639 of 40
Ray-intersection target with learned fusion+0.069738 of 40
Coefficient update+0.075240 of 40
Ground-truth normal offset+0.174240 of 40
Ground-truth 3D offset+0.176840 of 40

These rows come from Tables 1, 4 and 9.

  • The 3D head is ahead by +0.0054 c.u., which is 6% of the gain. With a point-to-plane fit the two heads tie, and the surface quality is practically the same.
  • The scalar keeps 98.5% of the gain of the full ground-truth 3D offset.
  • The surface target is ahead of the ray target by +0.0205 c.u., with a 95% interval of +0.0156 to +0.0257. This is nearly four times the difference between the two heads.

A vertex-displacement refiner of this kind is itself a surface-aligned correction, so this result supports the claim of the paper. What matters is that the evidence and the correction are expressed on the current surface, and the compact scalar is sufficient within that space. Therefore, beyond established surface-aligned features and fitting, the interface provides a correction space that is better than ray space and coefficient space under matched conditions, and one predictor can serve several face models and reconstructors.

Q2. How sensitive are the results for external reconstructors to the scan-assisted alignment?

Thank you for raising this point, which matters for practical use. The answer is that the results do not depend on the scan-based placement. Appendix F.4 and Table 17 report a new experiment in which all five starts of Table 2 are placed from the images alone. A rigid pose is fitted to landmarks detected in the same five calibrated views, and no scan information is used. The scoring applies no alignment to the output, so any placement error counts against the method.

  • Size of the placement error. The image-based placement differs from the scan-based one by 4.6° to 5.4° in median rotation and by 4.8 to 5.5 c.u. in median position.
  • Effect on our gains. They change by +0.017, +0.030 and +0.004 c.u. on DECA, EMOCA and Pixel3DMM, and the 95% intervals of these changes include zero. On MICA the gain is 0.058 c.u. lower, which is 4% of its gain of +1.47 c.u. On the landmark start it is 0.323 c.u. higher.
  • Effect on the ranking. Ours stays ahead of the photometric and landmark fit by +0.175 to +0.635 c.u., and ahead of landmark-only refinement by +0.719 to +1.088 c.u., on all five starts.

So the improvements carry over to a pipeline in which the start is placed without any reference scan, including pose errors of about 5°.

AuthorsOpenReview ↗

Response to Reviewer 8c4F (2/2)

W1. Validation under imperfect cameras, large pose errors and low-quality starts.

Thank you for this concern. In short, the method needs calibrated cameras, and within that setting it is robust. We address the four parts of your comment in turn.

  • Joint camera errors. Table 14 in Appendix E.3 now includes rows in which focal length, rotation and camera centre are perturbed together. At 1%, 0.1° and 1 c.u., the gain is +0.058 c.u. at α = 0.5 and +0.182 c.u. at α = 1. This error level moves the face by about 2.5 pixels in the 224-pixel crop, or about 20 pixels at full resolution, which is far above the reprojection error of a calibrated rig. At 5%, 0.25° and 2 c.u. the gain is still +0.038 c.u. at α = 1, and it turns negative at α = 0.5.
  • Uncalibrated cameras. The method is designed for calibrated capture, and Section 6 states this requirement together with the tolerance above. Cameras estimated by VGGT are about 50 crop pixels off and lie outside this range.
  • Large initial pose errors. The image-only placement in Q2 adds about 5° of rotation error and about 5 c.u. of position error, and the gains hold.
  • Low-quality starts. The starts in Table 2 are 5.3 to 5.7 c.u. from the scan, compared with 2.0 c.u. for the α = 0.5 start. From these starts our method gains +1.41 to +1.55 c.u. and improves 37 to 39 of the 40 identities, as Tables 3 and 16 show. The α = 2 start in Table 7 is twice as far from the scan fit as the image prediction. From it, one round gains +0.4125 c.u. and improves all 40 identities.

Regarding topological defects, the predictor works per point and accepts any mesh. The fit uses the topology of the face model, so a defective mesh would first be fitted to the model.

In summary, calibration is a real requirement of the method. Within the accuracy of calibrated rigs, the method is robust to joint camera error, to pose errors of about 5° and to starts that are far from the scan.

W2 and Q3. Is the cross-dataset gain a systematic correction or a subject-specific refinement, and what is the cause?

Thank you for asking for this analysis. On Headspace the gain is mainly a correction of Pixel3DMM's systematic error. Section 6 states this, and the revised abstract now says it as well. Appendix F.5 now analyzes the cause, and the FaceScape results show clear subject-specific refinement.

Why the shared part dominates on Headspace:

  • Pixel3DMM's own error there is mostly shared. Through the same fit, the mean ground-truth offset gives 0.255 mm, and each subject's own ground-truth offsets give 0.349 mm. So even perfect evidence is 73% shared.
  • The rig offers little multi-view evidence. Two of the five pods see 73% and 75% of the face, and the other three see under 5%.
  • Even so, Table 18 shows that a subject's own images give 0.030 mm more gain than another subject's images, with a 95% interval of 0.023 to 0.037 mm.

Subject-specific refinement on FaceScape, after the shared correction is removed:

  • From the bias-corrected Pixel3DMM start, the subject's own field beats the shared correction applied a second time by +0.208 c.u. The 95% interval is +0.163 to +0.253, and 37 of 40 identities improve.
  • By anatomical region, the own field exceeds the population-mean field by +0.193 c.u. on the chin, +0.166 c.u. on the lips, +0.086 c.u. on the nose and +0.071 c.u. on the eye region. Appendix E.2 reports this.
  • By expression, it exceeds the population-mean field on all 20 expressions, by +0.065 to +0.150 c.u., and every 95% interval is above zero.

Therefore, the cross-dataset result shows that a predictor trained on one dataset transfers a useful systematic correction to another without retraining, and this correction is about half of what the mean ground-truth correction achieves. Individual geometric refinement is demonstrated within FaceScape, across five reconstructors and all facial regions. Its transfer across capture rigs is weaker, and the paper states this boundary in Section 6.

Thank you again for the supportive and precise review. We hope the new experiments resolve your questions on novelty, placement and transfer, and we would be glad to provide any further detail.

AuthorsOpenReview ↗

Follow-up to Reviewer 8c4F

Dear Reviewer 8c4F,

Thank you again for your positive and careful review. We understand that this is a busy period, so we briefly summarize how our response addresses your questions.

  • Q1, a matched vertex-displacement refiner. It gains +0.0956 c.u. against +0.0902 c.u. for our scalar head, while the surface target is ahead of the ray target by +0.0205 c.u. Tables 4 and 9 report this.
  • Q2, scan-assisted alignment. With all starts placed from the images alone, our gains change by at most 0.06 c.u. on the four reconstructors, and the ranking of methods is unchanged. Table 17 reports this.
  • W1, imperfect cameras. Under joint errors of 1%, 0.1° and 1 c.u. in focal length, rotation and camera centre, the gain is +0.058 c.u. at α = 0.5 and +0.182 c.u. at α = 1. Table 14 reports this.
  • W2 and Q3, the cross-dataset gain. Appendix F.5 analyzes its cause. On FaceScape, the subject's own field beats the shared correction by +0.208 c.u.

We would be glad to answer any remaining question. If our response has resolved your concerns, we would be grateful if you could consider updating your assessment.

Thank you very much for your time.

Reviewer_6vhjOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper predicts a normal-direction correction at points on an existing face mesh from calibrated images, then fits a parametric face model to the corrected points. On FaceScape it reports a 4.5% reduction in symmetric Chamfer distance from a strong start. The controlled comparison of correction targets is a useful study, and the authors are unusually open about the settings where the gain shrinks.

What I am not sure about by far is why this two-stage pipeline is the right design rather than one of several viable ones, and how much of the result is due to the part that is genuinely new.

Critical strengths

The matched target comparison is a clean, informative study. Holding the encoder, decoder capacity and fitting stage fixed while varying the prediction target (Table 1, Figure 2, Appendix A) makes the comparison legible. However, the ray targets are constructed for this purpose rather than taken from published systems.

Limitations are reported rather than hidden. Weak-start counterexamples, camera sensitivity, the rougher free-form output and the small subject-specific margin on Headspace are all in the paper (Tables 7, 8, 10, 12), which makes the claims easier to assess.

Critical weaknesses
  • W1 — The paper does not yet make clear why the pipeline has to take this form, or what it buys over the alternatives it implies. The introduction explains what the pipeline is and how it is motivated by the choice of correction space, but I came away without a clear sense of the need being addressed: if the output must ultimately be parametric coefficients, it is not obvious what the surface prediction adds over refining the coefficients directly, and how large that addition is. The paper also never compares against pipelines that skip one of the two stages.

    • Observations: The predicted field looks more effective before fitting than after (0.13 c.u. vs 0.09 c.u. at α=0.5, Table 8), and the share of the available headroom that survives fitting falls as the start improves (0.34→0.28 with the bilinear fit, 0.30→0.19 with FLAME between α=0.5 and α=0.25, Table 11). The strongest controlled start gains about 0.037 c.u. (Table 7). Read together, these suggest the fitting stage, not the predicted evidence, is what limits the reported benefit, and that the benefit is concentrated on starts that are already weaker than the fitting model's own fit.
    • What would help: State the objective explicitly — the best mesh expressible by a fixed face model, or the best mesh overall — and then compare pipelines under matched input and supervision: fitted versus free-form output (including the normal-error and roughness cost of the free-form one), and an expressive single-image model such as monoNPHM as an additional start in Table 2. I could not find NPHM or monoNPHM anywhere in the paper; if such models were considered and set aside, a sentence explaining why would remove the question. Reporting the realized fraction for each downstream model would also show how the benefit scales with model capacity.
  • W2 — Set the framing aside, the remaining novelty is a fairly specific combination, and its margin over the closest controls is small. The method expresses a per-point correction in a local frame at the current mesh point, gives it a single signed degree of freedom along the current normal, and then fits a parametric model. Each ingredient has precedent: fusion of calibrated views in a shared 3D frame (TEMPEH, GRAPE), normal-direction detail and displacement-map style targets, point-to-plane residuals, and standard 3DMM fitting with per-coefficient weighting (§3.4). The authors do attribute several of these (§3, Eq.(1), Eq.(8)), so my concern is not about misattribution but about how large the increment is once the established parts are removed.

  • W3 — The dependence of the result on the shape of the starting error, and on repeating the correction, is not tested. Because the intermediate evidence is one scalar along the current normal, the reachable corrections are tied to the starting surface, so the outcome plausibly depends on which error the start carries, not only on how large it is.

    • Observations: The controlled starts vary the size of the coefficient error along a single line (Eq. 9) rather than its direction, so all reported starts share one error shape. The supplied implementation runs the correction once, whereas the coefficient-refiner control uses three steps (Appendix C.2). Table 2 uses different external starts, but their error size and placement are not controlled tightly enough to answer this question.
    • What would help: Compare starts with comparable Chamfer error but different error shapes, report how much of each start's error lies along its current normals, and test one additional pass with the same frozen predictor. A short discussion of whether the method is intended as a one-shot refinement or as an iterative one would also help readers place the result.

Minor Issues

  • Method and data specification. The bilinear model is used throughout but never described in the main text; spelling out that it is a FaceWarehouse-style bilinear identity × expression 3DMM (and whether BFM or FLAME could serve as well) would make the setup self-contained. I also could not find a single place that states the training data format: the start construction (Eq. 9), the target and correspondence rules (Appendix B), the training schedule (Appendix E.1) and the data split are spread across the paper. Similarly, the decoder's structure, inputs and outputs appear only in Appendix E.1, and "frozen" is used for two different things — the pretrained encoder and the decoder trained once on FaceScape and then transferred. Listing the required inputs (initial mesh and placement, calibrated images, all calibrated cameras, the parametric model, and the working-frame units) and how each is obtained would also help; the origin of the starting coefficients is not fully clear to me.
  • Coverage and visualization. It would help to state that points not seen in any selected view remain anchored even though they are inside the fitting support, and to report how many points are seen by how many views — greater coverage may explain part of the multi-view gain. On visualization, the error maps are informative but the gains are becoming small enough that numbers alone are hard to judge subjectively; a few rotating initial/refined comparisons and a face-relative scale for the canonical-unit errors would make the improvement easier to assess. I would not treat the absence of a user study by itself as a reason to reject.
  • Terminology and notation. "Aligned" carries several distinct meanings in the paper (the correction being surface-indexed, camera registration, similarity alignment for evaluation); a more specific term such as "surface-indexed" or "surface-anchored" for the first sense would avoid ambiguity. In §3.2, i is introduced in the prose while the displayed symbols only carry v. Finally, the topology statement would be clearer if stated in one sentence: prediction is topology-agnostic because it is a per-point function, while fitting relies on vertex-index correspondence, so there is no topology conversion step between the stages.
Questions for authors

Please refer to previous sections.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 6vhj (1/3)

Thank you very much for the constructive and detailed review. We appreciate your remarks on the matched-target study and on the open reporting of limitations. Your suggestions under W1 to W3 were concrete, and we followed them. Section, table and figure numbers below refer to the revised PDF, where the changes for your points are marked in blue and purple. Gains are Chamfer reductions on the FaceScape test set in canonical units, where 1 c.u. is about 0.82 mm.

W1. Why does the pipeline take this two-stage form, and what does it buy over the alternatives?

Thank you for asking us to state this explicitly. The objective is the best mesh that a fixed face model can express, because the reconstructors we refine and their downstream uses work with model coefficients. Section 3 now states this objective. Under this objective, the two-stage design is better than both single-stage alternatives, and the submission already contains both comparisons.

  • Without the surface stage. A coefficient refiner reads the same five-view evidence and updates the coefficients directly. Table 7 in Appendix C.2 shows that it is weaker at the three starts of the refinement setting. At α = 0.25 it loses 0.0278 c.u., while ours gains +0.0372 c.u. At α = 0.5 it gains +0.0752 c.u. against +0.0902 c.u. At α = 1 it gains +0.0414 c.u. against +0.2102 c.u. It is ahead of one round of ours only at α = 2, which is twice as far from the scan fit as the image prediction.
  • Without the fit. Table 12 in Appendix E.2 compares free-form and fitted output, including normal error and roughness. The free-form surface has a slightly lower Chamfer distance, 1.909 against 1.932 c.u. But it is about 30% rougher, 0.316 against 0.245, and its normal error is 23.37° against 21.62°. The fit keeps the surface quality of the start.

You observed that the field looks more effective before fitting than after, and that the fit limits the benefit. We agree that the face model bounds what any evidence can achieve, and this is intended under the objective above. However, the evidence decides how much of this bound is reached.

  • The bound applies to perfect evidence as well. Ground-truth offsets gain +0.1794 c.u. through the same fit, as Table 1 shows.
  • Within this bound, the realized fraction in Table 1 rises from 0.05 for the ray-projection target to 0.13 for one surface view and to 0.50 for five surface views.
  • With our main model, the fraction does not fall for stronger starts. Appendix C.5 reports with Table 10 that one round reaches 0.51, 0.52 and 0.56 of the ground-truth reduction at α = 0.25, 0.5 and 1. The values you quote from Table 15 were measured with the ResNet-50 encoder and a forward Chamfer metric on the region common to both models.
  • The benefit is not limited to weak starts. At the strongest start, α = 0.25, ours improves 38 of 40 identities.

You also asked for the realized fraction per face model. In Table 15 it is 0.34 for the bilinear model and 0.30 for FLAME at α = 0.5, and Section 5.3 now reports these values.

Regarding MonoNPHM, it is an implicit head model without a coefficient space. The fit has no parameters to refine there, so it is outside the scope of this method, and Section 6 now states this. Within this scope, Table 2 covers five starts, including Pixel3DMM as a recent and strong single-image method.

In summary, the two-stage form is the design that gives the best model-expressible mesh. The surface stage is where image evidence is easiest to predict, and the fit keeps the output a valid and smooth face-model mesh.

AuthorsOpenReview ↗

Response to Reviewer 6vhj (2/3)

W2. Once the established parts are set aside, how large is the increment?

Thank you for separating attribution from the size of the increment, which is a fair way to pose the question. The increment is the choice of the correction space, and the matched controls show that it is the dominant factor.

  • Coordinate system. With the encoder, decoder capacity, data and fit fixed, the surface target gains +0.0902 c.u., and a ray target with learned five-view fusion gains +0.0697 c.u. The difference is +0.0205 c.u., with a 95% interval of +0.0156 to +0.0257. Table 9 in Appendix C.4 reports this.
  • Coefficient space. A coefficient update from the same evidence gains +0.0752 c.u., as Table 1 shows.
  • Output head. A head of matched capacity that predicts a free 3D vertex offset gains +0.0956 c.u., which is +0.0054 c.u. more than our scalar head. Normal error and roughness are practically the same for both heads. Appendix B.1 reports this with Table 4.

You note that the ray targets were constructed for this study. This is deliberate. A matched comparison needs targets that differ only in the coordinate system, whereas published ray-space systems also differ in encoder, data and fitting.

So the same evidence yields +0.0697 c.u. along rays, +0.0752 c.u. in coefficient space and +0.0902 c.u. on the surface. At α = 0.5 the surface is ahead of these two controls by 29% and 20% of their gains. At α = 1 the gap is much larger, since Table 7 shows +0.0414 c.u. for the coefficient update and +0.2102 c.u. for the surface field. Therefore, what remains after setting the established parts aside is the correction space itself, and it accounts for a clear and consistent share of the gain.

W3. Does the result depend on the shape of the starting error and on repeating the correction?

Thank you for these two suggestions. We ran both, and Appendix C.5 reports them with Table 10 and Figure 5.

Repeating the correction with the same trained predictor helps.

  • From the bilinear starts, a second round adds +0.024 c.u. at α = 0.5 and +0.106 c.u. at α = 1. Three rounds reach +0.361 c.u. at α = 1. Section 5.2 reports this.
  • From the five reconstructor starts, the gain rises from +0.54–0.64 c.u. after one round to +1.15–1.30 c.u. after three rounds, and to +1.41–1.55 c.u. with a number of rounds chosen on validation. Tables 10 and 16 report this.
  • You noted that the coefficient refiner uses three steps while ours ran once. In Table 7, even one round of ours is ahead of the three-step refiner at α = 0.25, 0.5 and 1.

So the method is best used iteratively. One round is enough near the model's own fit, and distant starts benefit from more rounds. Section 3.4 now explains this.

The shape of the error matters as well.

  • As you asked, Table 10 reports how much of each start's error lies along its current normals. This share is 0.53, 0.62 and 0.74 for the bilinear starts at α = 0.25, 0.5 and 1, and 0.82 to 0.85 for the reconstructor starts.
  • To compare starts of equal size and different shape, we built a validation start with the same Chamfer error as α = 0.5, which is 2.031 c.u., but with error only in the expression coefficients. One round gains 0.0257 c.u. less from this start, with a 95% interval of 0.0139 to 0.0381. Two and three rounds show the same pattern.
  • The normal share of the two starts is almost equal. The expression-only start has a shorter spatial correlation length, by 2.9 c.u.

In summary, the gain depends on the spatial structure of the start error and not only on its size, which suggests that spatially coherent errors are the easiest to correct. This gives a practical indication of when the refinement pays off most.

AuthorsOpenReview ↗

Response to Reviewer 6vhj (3/3)

Minor issues

Thank you for these careful suggestions. All of them improved the paper, and each one is addressed in the revised PDF.

Method and data specification.

  • The bilinear model is FaceScape's own identity-by-expression model with 50 identity and 51 expression coefficients. Section 4 now describes it, together with the format of a training cell and the split into 280, 39 and 40 identities.
  • FLAME serves equally well as the fitting model. All results in Table 2 and on Headspace use FLAME, and Table 15 compares the two models directly. The fit in Eq. 8 only needs vertex positions as a function of the coefficients, so other linear models such as BFM fit the same formulation. Appendix E.1 now notes this.
  • Section 3 now lists the inputs in one place. They are the start mesh with its coefficients, the K calibrated images with their cameras, and the face model. No scan information is used at inference.
  • The starting coefficients come from two sources. On FaceScape they interpolate between the scan fit and the prediction of a ResNet-50 coefficient regressor, as in Eq. 9. For external reconstructors they are the reconstructor's FLAME coefficients plus a fixed per-vertex offset.
  • "Frozen" now refers only to the pretrained encoder, both in the text and in Figure 1. The decoder is called the trained predictor, and it is applied to other face models and datasets without retraining.
  • The decoder is a small MLP. The per-view network has layers of 1539, 256 and 256 units, and the decoder has layers of 518, 128 and 1. Eq. 5 and Eq. 6 in Section 3.3 give its inputs, and Appendix E.1 gives the full specification.

Coverage and visualization.

  • Points that no selected view sees remain anchored to the start. Section 3.4 now states this, and it also states that the support is the same for every K.
  • Section 4 now reports the coverage. On the test set, 14%, 25%, 22%, 21% and 17% of the fitted points are seen by one, two, three, four and five views. This is 3.0 views on average.
  • Coverage does not explain the multi-view gain. Averaging five one-view predictions uses the same views and the same support, and in Table 1 it gains +0.0509 c.u. against +0.0902 c.u. for the five-view predictor. Appendix E.1 now notes this as well.
  • Figure 6 in Appendix F.3 adds a qualitative comparison from three starts, with error maps and zoomed views of the nose, lips and chin. Section 4 now gives the scale, which is about 0.82 mm for 1 c.u.

Terminology and notation.

  • We keep "surface-aligned" for the correction, in line with the established terms pixel-aligned and mesh-aligned, and Section 3 now defines it once. The other two senses are now called registration.
  • Section 3 now introduces the index i together with the displayed symbols.
  • Section 3.4 now contains the single topology sentence you suggested.

So every minor point is resolved in the text, and none of them changes a reported number.

Thank you again for the detailed and helpful review. We hope the new experiments on the pipeline alternatives, on the size of the increment and on the error shape answer your main questions, and we would be glad to clarify anything further.

AuthorsOpenReview ↗

Follow-up to Reviewer 6vhj

Dear Reviewer 6vhj,

Thank you again for your constructive and detailed review. We understand that this is a busy period, so we briefly summarize how our response addresses your three main points.

  • W1, the two-stage design. Tables 7 and 12 compare with both single-stage alternatives. At α = 1 the coefficient refiner gains +0.0414 c.u. against +0.2102 c.u. for ours, and the free-form output is about 30% rougher than the fitted one.
  • W2, the increment. With everything else matched, the same evidence gains +0.0697 c.u. along rays, +0.0752 c.u. in coefficient space and +0.0902 c.u. on the surface.
  • W3, error shape and repeated correction. We ran both experiments you suggested, and Appendix C.5 reports them. A second round adds +0.106 c.u. at α = 1, and at equal start error the gain depends on the spatial structure of the error.

The minor points are all addressed in the revised PDF. We would be glad to answer any remaining question. If our response has resolved your concerns, we would be grateful if you could consider updating your assessment.

Thank you very much for your time.

Reviewer_MQ7uOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method for refining an existing 3D face reconstruction using multiple calibrated images. The central contribution is the representation used to predict the corrections. Rather than predicting corrections in image space (e.g. depth changes along camera rays) or directly updating the parameters of a 3DMM, the idea is to predict a scalar displacement along the surface normal at each point on the reconstructed mesh.

This has two claimed advantages: view independence (all cameras provide evidence about the same 3D surface point, making it natural to combine information across views) and model independence (the predicted corrections are geometric quantities rather than 3DMM coefficients, meaning the same predictor can be used with different face models and initial reconstruction methods).

Critical strengths

The paper identifies a sensible limitation of camera-ray-based correction: depth displacements become poorly conditioned at oblique viewing angles. The mathematical justification for using surface-normal displacements instead is straightforward and convincing.

Representing corrections in surface coordinates also provides a natural way to combine observations from multiple cameras while decoupling the learned predictor from the parameterisation of the downstream face model.

The authors conduct a substantial set of experiments designed to isolate the contributions of different components and include confidence intervals, multiple training seeds and identity-disjoint data splits. Importantly, they make a genuine effort to distinguish improvements due to their proposed representation from those arising from other aspects of the pipeline, even if the resulting evidence for the representation's significance is not especially compelling.

Critical weaknesses

The proposed method combines several well-established techniques: surface-aligned image features, multi-view feature aggregation, surface-normal displacement prediction and regularised parametric model fitting.

The key idea of the surface-normal displacement representation reminded me of the 2005 paper by Nehab on combining positions and normals. Although the setting is different there and the evidence arrives at the vertices from different sources, it shows others have thought along similar lines for two decades. In particular, Pixel2Mesh++ already demonstrated multi-view feature alignment on an existing mesh for geometric refinement, while DECA and other facial reconstruction methods have used learned displacement fields.

The main novelty therefore appears to be the specific combination of these components, rather than a fundamentally new representation or learning methodology. The paper does not adequately establish what distinguishes its contribution from these closely related approaches.

Although the overall reconstruction improvements are reasonable, the evidence supporting the specific methodological contribution is less compelling. The comparison with ray-based correction demonstrates improved residual prediction, but the corresponding improvement in reconstruction accuracy is small. For example, with one view, the Chamfer reduction improves from 0.0214 to 0.0238 when switching from ray-intersection to surface-normal correction. The substantially larger gains arise from incorporating additional calibrated views, something that is already well understood to benefit 3D reconstruction. Consequently, the experiments establish that the proposed representation works, but provide less convincing evidence that it constitutes a significant advance over existing approaches.

The cross-dataset results are particularly concerning. On Headspace, the proposed method reduces Chamfer error by 0.1194 mm, while a simple population-mean correction achieves 0.1163 mm. Thus, approximately 97% of the reported improvement can be reproduced without subject-specific image evidence. The difference between these two methods does not appear to be statistically significant. Furthermore, landmark-based refinement achieves statistically comparable final accuracy for Pixel3DMM and MICA initialisations on FaceScape. These findings weaken the claim that the method provides a generally useful, transferable mechanism for recovering subject-specific geometric errors.

The evaluation primarily compares alternative correction representations within the authors' own framework, alongside a coefficient-regression baseline and landmark-based refinement. However, the method is not compared directly against existing mesh-based, multi-view deformation approaches such as Pixel2Mesh++, or more recent multi-view facial reconstruction methods. Given that the proposed method requires calibrated multi-view imagery and an existing reconstruction, such comparisons are important for establishing whether it offers a meaningful advantage over established alternatives operating under similar conditions. This omission makes its practical significance difficult to assess.

Questions for authors

Given that previous work has already demonstrated multi-view feature aggregation at mesh vertices for learned geometric refinement (Pixel2Mesh++), and that surface displacement fields and normal-based corrections are established in facial reconstruction, what specifically is the new methodological insight beyond restricting the correction to a scalar normal displacement and subsequently fitting a parametric model?

Given the availability of multiple calibrated images and an initial reconstruction, how does the proposed learned refinement compare with conventional multi-view 3DMM fitting in terms of reconstruction accuracy, robustness and computational cost?

Does the method recover meaningful subject-specific geometric detail on Headspace, rather than predominantly correcting systematic biases in the initial reconstructions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer MQ7u (1/3)

Thank you very much for the thorough review. We appreciate that you found the geometric argument convincing and the experiments substantial, and we are grateful for the pointers to Pixel2Mesh++, DECA and Nehab et al. Your questions led directly to new experiments. Section, table and figure numbers below refer to the revised PDF, where the changes for your points are marked in red and purple. Gains are Chamfer reductions on the FaceScape test set in canonical units, where 1 c.u. is about 0.82 mm.

Q1. What is the new methodological insight beyond Pixel2Mesh++ and learned displacement fields?

Thank you for this question, which goes to the core of the paper. The insight is that the space in which a learned correction is expressed decides how well it can be predicted and combined across views, and that the current surface is the right space. The paper does not claim a new network component. It asks a question that earlier work left implicit, and it answers this question with matched comparisons.

The submission already contains evidence for this.

  • Section 3.1 and Eq. 2 show why a correction along a camera ray becomes poorly conditioned at oblique views.
  • Figure 2a confirms this on held-out identities. With the same encoder, decoder and data, the surface target exceeds the ray-intersection target by 0.16 in residual correlation at oblique views.
  • Table 1 shows that a coefficient update from the same five-view evidence gains +0.0752 c.u., while the surface field gains +0.0902 c.u.
  • Tables 1 and 2 show that the same trained predictor feeds the bilinear model and FLAME, and that it refines five different reconstructors without retraining.

Your question about Pixel2Mesh++ motivated a new experiment in Appendix B.1. We built its mechanism inside our pipeline, as a head of matched capacity that predicts a free 3D vertex offset from the same projected features. Table 4 shows that it gains +0.0956 c.u. against +0.0902 c.u. for our scalar head, a difference of +0.0054 c.u. Under a point-to-plane fit the two heads tie. Through the same fit, the ground-truth normal offset reaches 98.5% of the gain of the full 3D offset. In contrast, Table 9 in Appendix C.4 shows that moving the target from the camera ray to the surface changes the gain by +0.0205 c.u., which is nearly four times as much.

Regarding the three works you mention:

  • Pixel2Mesh++ deforms a template for generic objects and has no parametric model downstream. Our experiment shows that its vertex-offset head works as one instance of surface-aligned correction, and that the gain comes from the surface-aligned evidence that both heads share.
  • DECA's displacements are a single-image detail layer in UV space. Ours are multi-view residuals on an arbitrary current mesh, and a parametric fit consumes them.
  • Nehab et al. combine two measured quantities, positions and normals. In our method the normal defines the direction in which learned evidence is expressed. Section 2 now cites and discusses this work.

In summary, the new insight is the choice of the correction space. The matched experiments show that this choice has about four times the effect of the output head, and that it is what lets one predictor be reused across face models and reconstructors.

AuthorsOpenReview ↗

Response to Reviewer MQ7u (2/3)

W. The one-view gain over the ray target is small, and larger gains come from more views.

Thank you for the close reading of Table 1. The one-view numbers are correct. At one view the two targets differ only by the obliquity factor in Eq. 2, so the gap is expected to be small there, and the paper states that it is smallest for frontal views. However, the benefit of the surface target appears exactly where the analysis predicts it, and it grows when views are combined.

  • At oblique views the surface target is ahead by 0.16 in residual correlation. Figure 2a shows this.
  • Adding views without learned fusion is not enough. Table 1 shows that averaging five one-view predictions gains +0.0509 c.u., while the five-view surface predictor gains +0.0902 c.u. from the same images.
  • The new Table 9 in Appendix C.4 repeats the comparison with a ray model that also learns to fuse five views. It reaches +0.0697 c.u., against +0.0902 c.u. on the surface. The surface advantage grows from +0.0024 c.u. at one view to +0.0205 c.u. with fused views, and the 95% interval of the latter is +0.0156 to +0.0257.

So the gain from more views is not a separate effect that any representation would obtain. Views help much more when they share one surface reference, and this is the contribution of the representation.

W and Q2. Comparison with Pixel2Mesh++ and with conventional multi-view fitting, in accuracy, robustness and cost.

Thank you for asking for this comparison. It is now in Table 3 of Section 5.3, with full results in Table 16 of Appendix F.3. We compare with three calibrated multi-view fits that use the same five views and the same starts. They are the multi-view FLAME fit driven by Pixel3DMM's predictions, a photometric and landmark fit, and a FLAME fit to the VGGT reconstruction. The released Pixel2Mesh++ model is trained for generic object categories, so we compare with its mechanism in Table 4, as described under Q1.

  • Accuracy. Our method gains +1.41 to +1.55 c.u. across the five starts. The Pixel3DMM multi-view fit gains +0.96 to +1.33 c.u., and the photometric fit gains +0.86 to +1.17 c.u. Ours is ahead on every start, and every paired 95% interval against these two fits is above zero.
  • Robustness. Ours improves 37 to 39 of the 40 identities. The multi-view fits improve 31 to 38.
  • Cost. Ours takes 0.63 to 0.84 seconds per scan. The Pixel3DMM multi-view fit takes 7.95 to 12.4 seconds, so ours is about ten times faster.
  • Complementarity. One round of ours on top of the converged multi-view fit adds +0.239 to +0.372 c.u., and it improves on that fit for all 40 identities on every start.

These results use several predict-and-fit rounds, with the number of rounds chosen on validation. A single round, as in Table 2, gains less than a multi-view fit that runs to convergence. However, with a comparable iterative schedule our method is clearly ahead at a fraction of the cost. Figure 6 in Appendix F.3 shows a qualitative comparison.

Therefore, under the same views and starts, our method is more accurate, more robust and faster than conventional multi-view fitting, and it also improves the result of such a fit.

AuthorsOpenReview ↗

Response to Reviewer MQ7u (3/3)

W. Landmark-only refinement reaches comparable accuracy from Pixel3DMM and MICA.

Thank you for pointing to this result. It holds for a single round, which is what Table 2 reports. Even there, our method improves all 40 identities from every start, while landmark-only refinement improves 26 to 29. With several rounds the difference is large on all five starts. Table 16 shows that ours exceeds landmark-only refinement by +0.889 c.u. from MICA and by +0.986 c.u. from Pixel3DMM, with 95% intervals of +0.641 to +1.108 and +0.819 to +1.145. On the other three starts the margin is +1.08 to +1.16 c.u.

So the predicted field contains much more than landmarks provide, and a single round simply did not use all of it.

W and Q3. On Headspace, the population-mean correction achieves 97% of the gain. Does the method recover subject-specific detail there?

We appreciate this careful observation. On Headspace the gain is mainly a correction of Pixel3DMM's systematic error. Section 6 of the submission states this, and the revised abstract now says it as well. However, this does not mean that the method only corrects bias, for three reasons.

  • The transfer is useful in itself. A predictor trained only on FaceScape removes 0.12 mm of error on a new dataset with a different capture rig, and it improves 698 of 776 subjects. Table 18 shows that this is about half of what the mean ground-truth correction achieves, although our predictor uses no Headspace ground truth.
  • Subject-specific signal is present on Headspace. Table 18 shows that a subject's own images give 0.030 mm more gain than another subject's images, with a 95% interval of 0.023 to 0.037 mm.
  • On FaceScape the gain is subject-specific. In Table 1 the population-mean field gains +0.0021 c.u. and another subject's images give −0.0059 c.u., against +0.0902 c.u. for the subject's own field. From the Pixel3DMM start, the subject's own field beats the shared correction applied a second time by +0.208 c.u., and 37 of 40 identities improve. Appendix E.2 further shows that the own field exceeds the population-mean field in every face region, most clearly on the chin and the lips.

Appendix F.5 also explains why the shared part dominates on Headspace. Even perfect evidence is 73% shared there, because the mean ground-truth offset gives 0.255 mm and each subject's own ground-truth offsets give 0.349 mm. In addition, only two of the five camera pods see the face.

In summary, the Headspace result shows that the predictor transfers a useful correction to a new dataset without retraining, and the FaceScape results show that it recovers subject-specific geometry. Subject-specific transfer across capture rigs is weaker, and the paper states this boundary in Section 6.

Thank you again for the detailed comments. We hope the new comparisons address your concerns about novelty and significance, and we would be happy to clarify anything further.

AuthorsOpenReview ↗

Follow-up to Reviewer MQ7u

Dear Reviewer MQ7u,

Thank you again for your thorough review. We understand that this is a busy period, so we briefly summarize how our response addresses your main points.

  • Novelty. A matched Pixel2Mesh++-style vertex-offset head differs from our scalar head by only +0.0054 c.u., while moving the target from the camera ray to the surface changes the gain by +0.0205 c.u. Tables 4 and 9 report this.
  • Multi-view fitting. Ours gains +1.41 to +1.55 c.u., against +0.96 to +1.33 c.u. for the Pixel3DMM multi-view fit, and it is about ten times faster. Table 3 reports this.
  • Subject-specific detail. On FaceScape, the subject's own field beats the shared correction by +0.208 c.u., and 37 of 40 identities improve.

We would be glad to answer any remaining question. If our response has resolved your concerns, we would be grateful if you could consider updating your assessment.

Thank you very much for your time.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General response to all reviewers (1/2)

We sincerely thank all three reviewers for their time and for their careful and constructive comments. We are encouraged that the reviews found the geometric argument convincing, the matched-target study clean and informative, and the controls rigorous. Your questions helped us strengthen the paper. We answer each of them below with evidence from the submission and with new experiments.

A revised PDF is uploaded. Changed text is coloured by reviewer: red for MQ7u, blue for 6vhj, green for 8c4F, and purple for points raised by several reviewers. All section, table and figure numbers in our replies refer to this revised PDF. Gains are reductions in face-region symmetric Chamfer distance on the FaceScape test set, averaged over 40 identities, in canonical units where 1 c.u. is about 0.82 mm.

Three questions were shared across the reviews. We summarize our answers here and give the details in the individual replies.

1. How does the method compare with conventional multi-view fitting?

In short, our method is more accurate than calibrated multi-view fitting, about ten times faster, and it still helps when applied on top of such a fit. Table 3 in Section 5.3 and Table 16 in Appendix F.3 report this comparison. All methods use the same five calibrated views, the same five starts and the same metric.

Methodlandmark fitDECAEMOCAMICAPixel3DMM
Pixel3DMM multi-view fit+1.049+0.957+0.961+1.135+1.331
Photometric and landmark fit+0.868+0.860+0.878+1.171+1.054
VGGT reconstruction and FLAME fit+0.080+0.339+0.353+0.768+0.586
Ours, iterated+1.424+1.411+1.546+1.471+1.480
  • Accuracy. Ours is ahead of the Pixel3DMM multi-view fit by +0.149 to +0.586 c.u. and ahead of the photometric fit by +0.301 to +0.669 c.u. Every paired 95% interval is above zero on all five starts.
  • Reliability. Ours improves 37 to 39 of the 40 identities. The multi-view fits improve 31 to 38.
  • Cost. Ours takes 0.63 to 0.84 seconds per scan. The Pixel3DMM multi-view fit takes 7.95 to 12.4 seconds.
  • Complementarity. One round of ours on top of the converged multi-view fit adds +0.239 to +0.372 c.u., and it improves on that fit for all 40 identities on every start.

Therefore, surface-aligned evidence is more than competitive with established multi-view fitting. It carries information that those fits do not extract. This holds although Pixel3DMM was trained on FaceScape, while our predictor never saw a validation or test identity.

2. What is new beyond established components?

Our contribution is the choice of the space in which a learned correction is expressed, together with controlled evidence that this choice matters. To test this directly, we built the Pixel2Mesh++-style alternative inside our pipeline. It uses the same encoder, projection, fusion and fit, and a head of matched capacity that predicts a free 3D vertex offset. Table 4 in Appendix B.1 and Table 9 in Appendix C.4 give the results.

  • Moving the target from the camera ray to the surface, with everything else fixed, raises the gain from +0.0697 to +0.0902 c.u. The difference of +0.0205 c.u. has a 95% interval of +0.0156 to +0.0257.
  • Replacing our scalar head by a free 3D offset head changes the gain by only +0.0054 c.u.
  • Through the same fit, the ground-truth normal offset reaches 98.5% of the gain of the full ground-truth 3D offset.

So the coordinate system has nearly four times the effect of the output head, and the compact scalar target loses almost nothing. This is the methodological point of the paper. The other components are kept standard on purpose, so that the comparison isolates this choice.

AuthorsOpenReview ↗

General response to all reviewers (2/2)

3. What does the Headspace transfer show?

The predictor was trained only on FaceScape. On Headspace it reduces the error of Pixel3DMM by 0.12 mm and improves 698 of 776 subjects, without any Headspace ground truth. Table 18 in Appendix F.5 shows that this is about half of the 0.25 mm that the mean ground-truth correction achieves through the same fit.

The reviewers are right that this gain is mainly a correction of Pixel3DMM's systematic error. Section 6 of the submission states this, and the revised abstract now states it as well. However, two results show that the representation itself carries subject-specific information.

  • On Headspace, a subject's own images give 0.030 mm more gain than another subject's images, with a 95% interval of 0.023 to 0.037 mm. Table 18 reports this contrast.
  • On FaceScape, from the bias-corrected Pixel3DMM start, the subject's own field beats the shared correction applied a second time by +0.208 c.u. The 95% interval is +0.163 to +0.253, and 37 of 40 identities improve. Section 5.3 reports this result.

Appendix F.5 also explains why the shared part dominates on Headspace. Even with perfect evidence, 73% of the achievable gain there is shared across subjects. The mean ground-truth offset gives 0.255 mm, and each subject's own ground-truth offsets give 0.349 mm. In addition, only two of the five camera pods see the face.

In summary, the cross-dataset transfer delivers a useful correction without retraining, and subject-specific refinement is demonstrated within FaceScape across five reconstructors.

4. Additional robustness results

  • Placement without the scan, asked by 8c4F. When every external start is placed from the images alone, our gains change by at most 0.06 c.u. on the four reconstructors, and the ranking of methods is the same on all five starts. Appendix F.4 and Table 17 report this.
  • One round or several, asked by 6vhj. A second round adds +0.024 c.u. at α = 0.5 and +0.106 c.u. at α = 1. From the five reconstructor starts, iteration lifts the gain from +0.54–0.64 c.u. after one round to +1.41–1.55 c.u. Section 5.2 and Appendix F.3 report this.
  • Shape of the start error, asked by 6vhj. At equal start error, a start with expression-only error is harder to correct. So the spatial structure of the error matters, and not only its size. Appendix C.5 reports this.
  • Joint camera error, asked by 8c4F. With focal length, rotation and centre errors applied together at 1%, 0.1° and 1 c.u., the gain remains positive at α = 0.5 and at α = 1. Table 14 in Appendix E.3 reports this.

Together, these results show that the gains do not depend on scan-based placement, on a single round, or on perfectly exact cameras.

Changes in the revised PDF

  • Section 5.3 has a new comparison with calibrated multi-view fitting in Table 3. Appendix F.3 gives the full results and a qualitative comparison in Figure 6.
  • New Appendix sections B.1, C.4, C.5 and F.4 report the new experiments with full protocols. Appendix E.3 has new rows for joint camera error.
  • Section 3 now lists the inputs and the objective, defines "surface-aligned" once, and explains iterative use.
  • Section 4 now describes the bilinear model, the training data format and the view coverage.
  • The abstract and Section 5.3 now state clearly what the Headspace transfer shows.

We hope these results address your concerns, and we would be glad to answer any further questions.

Area_Chair_CBKZOpenReview ↗

Invitation to Participate in the Discussion

The authors have provided their responses to the reviewers' comments. We would appreciate it if the reviewers could review the responses and share any further thoughts during the discussion period.

AuthorsOpenReview ↗

Reply to the Area Chair

Dear Area Chair,

Thank you very much for handling our submission and for inviting the reviewers to the discussion. For convenience, we list where each main concern is answered in the revised PDF.

  • Novelty, raised by all three reviewers. Tables 4 and 9, summarized in point 2 of the general response.
  • Comparison with multi-view fitting, raised by MQ7u. Table 3 in Section 5.3.
  • Design rationale and error shape, raised by 6vhj. Tables 7, 10 and 12.
  • Scan-free placement and camera error, raised by 8c4F. Tables 17 and 14.
  • The Headspace transfer, raised by MQ7u and 8c4F. Appendix F.5, summarized in point 3 of the general response.

We remain available for any further questions during the discussion period.

Thank you again for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.