acceptodds
Under review as a conference paper at ICLR 2027

G-VR: INVERTIBLE GEOMETRIC VOLUME REPRESEN- TATION FOR LEARNING

Abstract

Investigating geometric structure representation is key to explainable and controllable learning. We introduce G-VR, an invertible geometric representation for 3D volume data that enables lossless reversible encoding and exact restoration. Our framework operates on a standard 3D uniform grid, where the associated scalar field, whether voxel intensities from medical scans or occupancy ratios from solid geometries, serves as the target measure. Grounded in optimal transport and quasiconformal mappings on three-manifolds, G-VR encodes each volume into a canonical geometric signature: a quasiconformal deformation field that captures the intrinsic geometric structure of the data. Together with the extrinsic total mass of the volume, this representation guarantees exact restoration of both the geometry and the original scalar field. The representation naturally preserves geometric structure under interpolation and manipulation, while ensuring full invertibility. Experiments on both image-based and geometry-based volumes validate its fidelity, reversibility, and effectiveness in interpolation.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
3D volumetric representation, semi-discrete optimal transport, quasiconformal mapping, geometric representation, finite element methods, medical image reconstruction, volumetric data compression

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LaP3OpenReview ↗
Rating
2
Confidence
1
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces G-VR, an invertible geometric representation for 3D volume data that encodes volumes as quasiconformal deformation fields derived from semi-discrete optimal transport. Each volume is represented by a nodal 3D QC tensor field (97 vertices × 6 parameters) plus a scalar total mass, and can be restored via a cold FEM decoder using mass conservation. The same frozen protocol is applied uniformly to both image-based volumes (MRI, CT) and geometry-based volumes (tetrahedral meshes). Experiments demonstrate near-lossless restoration (SSIM=0.998 on MRI) and improved latent interpolation when G-VR features guide a small autoencoder compared to a voxel-only baseline.

Critical strengths
  1. Elegant mathematical formulation with a unified treatment across modalities. The idea of representing a volume by "how to deform a uniform distribution into the target density" is conceptually clean, and the combination of optimal transport with quasiconformal encoding provides theoretical invertibility guarantees (positive Jacobian, mass conservation). The same frozen protocol working across MRI, CT, and geometry volumes without modality-specific adaptation is a genuine advantage over methods that require separate pipelines.

  2. Extreme compression with near-lossless fidelity. Encoding 262,144 voxel values into 583 parameters (~450× compression) while achieving SSIM=0.998 is impressive. If this compression ratio holds at larger scales, it could be practically valuable for storage and transmission of volumetric data.

Critical weaknesses
  1. No external baselines — the experimental evaluation is entirely self-referential. The paper compares only its own two variants (Voxel-only AE vs G-VR-guided Hybrid). There is no comparison with any existing method: no volume compression approaches (JPEG2000-3D, wavelet-based, neural compression), no invertible representations (normalizing flows, INNs), no deformation-based methods (VoxelMorph, LDDMM, SyN), and no geometric representations (DeepSDF, Occupancy Networks). Without external references, it is impossible to assess whether G-VR offers any advantage over established alternatives, or whether the claimed compression-fidelity tradeoff is competitive.

  2. Minimal connection to learning, weak fit for ICLR. Despite the title "...for Learning," the core method is a purely mathematical pipeline (OT + QC + FEM) with no learned components. The only learning experiment is a small autoencoder (112K parameters) used for interpolation on 5 MRI pairs. The paper demonstrates no downstream learning task (classification, segmentation, generation) where G-VR provides a benefit. Perhaps this paper is not necessarily well suited for submission to ICLR.

Questions for authors
  1. The current template uses only 97 vertices at 64³ resolution. How does the method scale — if resolution increases to 128³ or 256³, how many vertices are needed to maintain SSIM>0.99? What is the computational cost (encoding time, decoding time) as a function of resolution?
  2. The paper claims the representation is useful "for learning," but no downstream learning task is evaluated. Can you demonstrate a concrete scenario where using G-VR as input representation improves a learning outcome (e.g., classification accuracy, segmentation Dice, or generation quality) compared to using raw voxels or other representations?
1 comment
AuthorsOpenReview ↗

## Response to Reviewer LaP3

We thank the reviewer for pointing out areas regarding template scale, external comparisons, and downstream learning. We would like to clarify that the computational scaling is well-managed via our newly introduced BFM, and the downstream gains directly stem from the topology-preserving nature of our geometric representation.

Q1: How many vertices are needed for 128³ volumes, and how does the computational cost scale with resolution?

Reply: In the original manuscript, nn denotes the number of vertices per axis. For 128³, BFM requires 7.19M vertices and 42.47M tetrahedra. By introducing BFM to replace SDOT, the actual computational cost is well-controlled.

Details:

  • Scale Comparison: 64³ (n=97n=97) has 0.91M vertices and 5.31M tetrahedra; 128³ BFM (n=193n=193) has 7.19M vertices and 42.47M tetrahedra. Halving the grid spacing increases elements by 8×.
  • Measured Cost: The main computational stages for 10 cases at 128³ using BFM take approximately 34.4 minutes in total, achieving SSIM 0.992067 and RMSE 0.003719. (Detailed solver configurations are now included in the revised Section 3.3 and Appendix G.)

Q2/W2: Can G-VR improve real downstream learning tasks?

Reply: The improvement is most pronounced in MRI reconstruction, with CT segmentation presented as exploratory evidence. Topology preservation plays a key role by eliminating artifacts in latent space interpolation.

Details:

  • Quantitative Gains:
Task & MetricWithout G-VRWith G-VRChange
CT 4-organ Seg. Dice0.78900.792118+0.003118
MRI AE SSIM0.936360.96610+0.02974
MRI Transformer AE SSIM0.9718190.972504+0.000685
  • Role of Topology Preservation: QC positive definiteness equals zero folding, eliminating topological artifacts. In Table 3, path second-difference RMS is reduced by 42.5%, intermediate frame RMSE by 9.7%, and local out-of-range rate by 77.7%. With matched latent payloads and <2% parameter difference, improvements are attributable to the geometric branch.
  • CT Segmentation: Evaluated with 3 random seeds, the patient-level 95% CI is [−0.000515, 0.017948] (includes zero). (The revised manuscript explicitly labels this as "exploratory evidence" in Section 4.4 and Appendix D to strictly limit our claims.)

Q3/W1: What advantages have been established, and how does it compare to external methods?

Reply: The core advantage of G-VR is providing an operable and learnable geometric structure representation, rather than universally leading in pure storage efficiency.

Details:

  • Storage Comparison: Compressed QC maintains high restoration quality (SSIM 0.998439), but Voxel+ZFP performs better in file size and error (see Appendix E).
  • External Learning Comparison: We adapted Dora-VAE on 3 geometric cases (Dice 0.765/0.724). Given vast differences in tasks and protocols, we strictly limit learning claims to settings supported by direct controls (like MRI AE). (We have added these external comparisons and the failure case analysis to Appendix F in the revised manuscript.)

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

The paper proposes G-VR, a representation that encodes a 3D scalar volume (MRI/CT intensities, or occupancy ratios of CAD solids) as a geometric deformation of a uniform reference domain. The volume’s normalized density is used as the target of a semi-discrete optimal transport map from the uniform cube. That map is interpolated onto a fixed 97-vertex tetrahedral template and encoded as per-vertex 3D quasiconformal (Beltrami) tensors, plus a scalar total mass. The tensors are then refined so that decoding matches the input. Decoding solves three elliptic FEM problems to recover the map and reads each voxel’s density from the inverse Jacobian of its containing tetrahedron. The paper claims an invertible, lossless, modality-agnostic, and structure-preserving representation. For validation, it reports near-lossless restoration on 100 OASIS MRI, 43 CT-ORG, and 40 SimJEB volumes at 64³, with all mappings orientation-preserving. It also shows that a voxel autoencoder augmented with G-VR features improves reconstruction and latent interpolation over a voxel-only autoencoder of matched size.

Critical strengths
  1. Principled mathematical construction. The representation is built on well-established theory: optimal transport for the canonical map, mass conservation for density recovery, and the 3D quasiconformal tensor with an elliptic decoder for map recovery. It also includes explicit validity checks (positive Jacobians, no flipped elements, tensor consistency). This gives the encoding a clear geometric interpretation and well-defined invertibility conditions in the continuous setting, which is more interpretable than learned latent codes.

  2. Unified, fixed-size, corresponding representation across modalities. Because every volume is a deformation of the same uniform template, image volumes (MRI, CT) and geometry volumes (CAD occupancy) map into the same fixed-length code with vertex-level correspondence. This is a useful property for learning and interpolation. The matched-capacity experiment offers initial evidence that these geometric features help an autoencoder produce smoother interpolations.

Critical weaknesses
  1. The “lossless” claim is not credible as described. G(V) consists of about 583 numbers (97 vertices × 6 tensor entries + 1 mass), while each volume has 64³ = 262,144 values. No representation of this size can be invertible for general volumes; a random-noise volume is an obvious counterexample. Moreover, the P0 decoder assigns one constant density per deformed tetrahedron, so a restored volume contains at most a few hundred distinct values. That is hard to reconcile with the reported near-perfect MRI restoration (3D-SSIM 0.998, RMSE 0.003). The encoder itself is also a per-volume optimization against the ground truth, with “no claim of global optimality.” Since invertibility is the headline contribution, this inconsistency undermines the paper’s central claim. The authors should clarify the actual template size, the serialized code size, and how restoration achieves this fidelity.

  2. Novelty is limited, and the evaluation is weak. The method is largely a 3D extension of G-IR (Chen et al., AAAI 2026), which uses the same OT → Beltrami-coefficient → restoration and interpolation pipeline in 2D, with Chen & Lui’s (2025) 3D quasiconformal representation in place of the 2D Beltrami coefficient. G-IR is mentioned only in passing, and what is new relative to it is never stated. The only comparison is against a voxel-only autoencoder. There are no baselines of comparable size, such as downsampled volumes, OT bucket masses, or vertex displacement fields, which would show whether the 6-number tensor form is better than simply storing 3D vertex positions. The hybrid autoencoder also receives G-VR features of the very volume it reconstructs, so its improvement is unsurprising and does not isolate the representation’s value. Finally, undefined jargon (“strict gate”, “candidate-quality gate”, “mapping-safe”, “cold decoding”) makes several reported results impossible to interpret.

Questions for authors
  1. How many tetrahedra does the template contain, and what is the serialized size of G(V) in bytes compared to the 64³ input? How can a code of about 583 values, decoded as piecewise-constant density per tetrahedron, achieve 3D-SSIM 0.998 on MRI? Is any per-sample information besides (B, m_V) used during restoration?

  2. The OT map does not necessarily send each face of the cube onto itself, yet the decoder’s boundary conditions assume it does. How is this reconciled, and how much restoration error does it introduce?

  3. What is new relative to G-IR (Chen et al., AAAI 2026) beyond the extension from 2D to 3D?

  4. Please define “strict gate”, “candidate-quality gate”, and “mapping-safe”, and explain why 6/43 CT-ORG and 2/40 SimJEB cases fail them.

  5. How long does encoding take per volume (SDOT, lift, and refinement), and how does it scale with resolution?

1 comment
AuthorsOpenReview ↗

Response to Reviewer UPXq

We thank the reviewer for raising key questions regarding the mesh scale, finite-dimensional encoding, boundary conditions, and failure criteria. We would like to clarify that the high SSIM achieved by the P0 decoder is driven by the accurate local volume changes captured by the refined QC field, rather than the polynomial order itself.

Q1: What is actually stored, and how does P0 achieve high SSIM?

Reply: n=97n=97 means 97 vertices per axis, totaling 912,673 vertices. Each vertex stores 6 independent QC components (5,476,038 scalars). P0 means constant density within each tetrahedron. High SSIM is primarily due to the QC field and node refinement ensuring accurate local volume changes, rather than the order of P0.

Details: The template contains ~5.31M tetrahedra, far exceeding the voxel count of the 64³ grid. Thus, query resolution is determined by tetrahedral size. Decoding uses Eq. 7 to recover values based on local volume changes and total mass. The independent decoding process uses only code values, shared template, and decoder. The average cropped SSIM for 20 MRI cases reaches 0.998440.

Q2: How does the OT map satisfy the decoder's boundary conditions?

Reply: During encoding, we apply boundary compatibility correction to the initial OT map (normal displacement is zero, tangential sliding is preserved). Decoding uses mixed boundary conditions on a per-face basis.

Details:

  • Correction Effect: Maximum normal offset before correction is 0.014–0.019, becoming zero after correction.
  • Importance of Tangential Sliding: With a fixed QC field, allowing boundary sliding yields SSIM 0.998440, whereas full fixation drops it to 0.528400.
  • Clarification on Optimality: Since boundary correction alters the initial OT map, the revised manuscript refers to it as a "boundary-compatibility corrected OT map" and does not claim it as an exact OT optimal solution. (Updated in Section 3.3.)

Q3/W2: What is new compared to G-IR, and what is the component evidence?

Reply: As addressed in 45pE Q2/Q3, G-VR reconstructs the complete restoration pipeline for 3D volumes.

Details: Node refinement is the core for maintaining high fidelity (SSIM drops to 0.494628 upon ablation). The map anchor trades off accuracy and geometric consistency (SSIM 0.999718 but inconsistency 0.103885 without it). All 12 independent decodes passed geometric checks for zero folding and zero solver failures. (See Appendix C.1 for full details.)

Q4: How are the thresholds defined, and why did CT-ORG and SimJEB cases fail?

Reply: Thresholds are divided into geometric validity (zero negative Jacobians, min detJ≥0.05, boundary checks) and restoration quality (specific SSIM/Dice thresholds). Failed cases all failed on quality metrics.

Details:

  • SimJEB: Cases 136 and 142 had Surface Dice@1 cell of 0.968485 and 0.985140 (<0.99), failing the strict gate, but Dice and HD95 were acceptable and mapping-safe.
  • CT-ORG: Case 000 had SSIM 0.807998 and RMSE 0.046917 (stalled after 12 consecutive step rejections); Cases 002, 005, 017, 018, 025 had SSIM <0.99 but acceptable RMSE. All 6 passed cold-valid. (Detailed analysis of these failure cases is now provided in Appendix F.2.)

Q5: What is the encoding time, and how does the cost scale with resolution?

Reply: The encoding time heavily depends on the OT solver. The original 64³ SDOT pipeline (per-vertex) takes 92.67–141.02 minutes, whereas the 128³ BFM pipeline requires only ~34.4 minutes, handling an 8× increase in tetrahedra with better fidelity.

Details:

  • Efficiency Comparison:
Resolution / MethodVertices/AxisTotal VerticesEncoding TimeAvg. SSIM
64³ / SDOT (per-vertex)97912,67392.67–141.02 min0.998440
64³ / SDOT (per-tetra)97912,6738.58–12.04 min0.998440
128³ / BFM1937,189,057~34.4 min0.992067
  • Detailed 64³ Timing (SDOT, 10 cases per class):
DataPer-tetrahedron QC Avg. (Range), minPer-vertex QC Range, min
MRI8.70 (8.66–8.72)114.33–115.76
CT10.78 (8.58–11.92)119.94–122.20
SimJEB9.85 (8.72–12.04)92.67–141.02
  • Storage Compression: Per-vertex storage reduces the raw payload from ~127.40 MB to 21.90 MB, an ~82.8% reduction. (The full timing table is now included in Appendix G.)
Reviewer_45pEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

G-VR represents a nonnegative scalar volume through a geometric deformation and its total mass. MRI/CT intensities and geometric occupancy values are converted into a common density representation. Semi-discrete optimal transport initializes a deformation of a shared tetrahedral template, which is encoded as a nodal quasiconformal (QC) tensor field and refined to reduce restoration error. Decoding solves finite-element equations to recover the deformation and uses mass conservation to reconstruct the scalar field. Experiments on MRI, CT, and geometry data report high restoration fidelity. A G-VR-guided hybrid autoencoder also improves reconstruction and latent interpolation over a voxel-only autoencoder.

Critical strengths

The representation has a clear geometric interpretation: the local tensors are derived from spatial deformation, and decoding is based on mass conservation.

The same pipeline handles different types of volumetric data. MRI, CT, and geometric occupancy fields enter the encoder through a common scalar-field interface, suggesting potential as a representation for 3D generation or reconstruction.

The experiments report high-fidelity restoration results. The comparison between the Hybrid and voxel-only AE provides preliminary evidence that G-VR can serve as a complementary module for autoencoders.

Critical weaknesses
  • The additional technical contribution over prior work is not sufficiently established. The overall approach follows the OT–QC image representation in G-IR, while 3D QC representations and PDE-based recovery already have been explored in previous papers. The novelty of the added nodal tensor refinement, shared boundary treatment, and decoding pipeline needs to be more clearly established.

  • The validation with stronger AE architectures is missing. The voxel-only AE versus Hybrid comparison supports the value of G-VR, but comparison only with a traditional voxel AE is insufficient to establish its advantages over newer 3D representations and architectures. For geometry reconstruction, the paper should compare reconstruction accuracy and computational cost against stronger AE/VAE models with recently proposed transformers, and evaluate whether G-VR's gains extend to these architectures.

  • Ablations of the core components are insufficient. The existing ablation mainly evaluates adding G-VR as a whole to the voxel-only AE, without isolating the contributions of individual encoding and decoding choices.

  • The experiments are largely limited to restoration and interpolation on small datasets at low resolution. The representation's expressive capacity, generalization, and computational scalability on larger and more diverse 3D datasets remain insufficiently validated.

Questions for authors
  • Why convert the OT mapping into QC tensors and then recover the mapping, rather than directly storing the OT mapping or displacement field?

  • How does the method perform, and what are its computational costs, on larger datasets and higher-resolution 3D volumes? To further support its potential for generative modeling, could the authors additionally provide a 3D generation experiment?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 45pE

We thank the reviewer for the detailed questions regarding the method's explanation, experimental validation, and scope. We appreciate the opportunity to clarify that G-VR is not merely an invertible autoencoder, but rather a structural upgrade from "scalar representations" to "deformation structure representations" for volumetric data.

Q1: Why encode the OT map into a QC tensor, and then recover the map from QC?

Reply: Storing the displacement field directly can also achieve restoration, but it is an unconstrained vector field: it does not guarantee bijectivity after interpolation and lacks local geometric meaning as a learning feature. The QC tensor (B=det⁡(J)(J⊤J)−1B=\det(J)(J^\top J)^{-1}) is a local metric of the deformation map, transforming the "deformation structure" into a representation with invertibility, interpolability, and learnability.

Details:

  • Theoretical Guarantee for Topology Preservation: BB is symmetric positive definite if and only if JJ is orientation-preserving and non-singular. During node refinement, we constrain λmin⁡(Bv)≥η\lambda_{\min}(B_v)\ge\eta and reject candidates with non-positive Jacobians (Eq. 5), theoretically eliminating the possibility of folding.
  • Storage vs. Restoration Quality (Average of 20 64³ MRI cases):
RepresentationAvg. Full File SizeCropped SSIM
QC, FP3220,021,264 B0.998440
QC + ZFP, rate=42,511,037 B0.993617
Direct displacement, FP329,097,180 B0.998440
Voxel + ZFP, atol=0.01100,297 B0.999997

Although displacement fields are smaller in raw volume, and voxel compression is superior in pure storage efficiency, QC compressed at rate=4 (2.51 MB) still maintains an SSIM of 0.993617 with zero folding. We choose QC because it provides geometric properties—interpolability, positive definiteness, and learnability—that other representations lack. (We have added the full storage comparison table and this discussion to Appendix E in the revised manuscript.)

Q2/W1: What is new compared to G-IR and 3DQC?

Reply: G-VR extends geometric deformation representation to 3D volumetric data. While adopting the OT-QC methodological framework, the target shifts from 2D images to 3D volumes, necessitating a complete reconstruction of discretization, constraints, and the restorationpipeline, alongside a more efficient OT solver.

Details:

  • Unified Measure and Pipeline Reconstruction: 3D volumes lack natural boundaries. We unify the scalar measure via standardization (Eq. 2) and reconstruct map initialization on a tetrahedral mesh.
  • Node Refinement for Volumetric Restoration: 3DQC addresses "how to represent a given map," whereas volume restoration requires "reversing the map to recover the scalar field." We designed a P0 reconstruction error (Eq. 5) to optimize nodal QC within FEM. Ablations show removing node refinement drops SSIM from 0.998440 to 0.494628—a core restoration mechanism absent in prior pipelines.
  • OT Solver Upgrade: For higher resolutions, we introduced a continuous optimal transport solver based on the Back-and-Forth Method (BFM) in 128³ experiments. BFM requires 193 vertices/axis and ~34.4 minutes, achieving SSIM 0.992067, outperforming SDOT (257 vertices/axis, SSIM 0.966067).

(We have updated Section 3.3 to detail the integration of BFM and added the full ablation study to Appendix C.1.)

Q3/W3: What are the roles of the core components?

Reply: Node refinement is the core for ensuring high-fidelity restoration; the map anchor trades off between accuracy and geometric consistency; the coefficient anchor has minimal impact under current weights.

Details:

  • Ablation Results (Average of 20 64³ MRI cases):
ConfigurationCropped SSIMTensor Inconsistency
Full configuration0.9984400.077549
Without map anchor0.9997180.103885
Without coefficient anchor0.9984730.077767
Without node refinement0.494628—

Removing the map anchor slightly increases SSIM but worsens tensor inconsistency (0.077549 → 0.103885), indicating its role in constraining geometric consistency. The coefficient anchor has negligible impact. Anchors are optimization terms, while positive definiteness, Jacobian, and boundary checks are hard validity criteria. All 12 independent decodes passed these checks with zero folding and zero solver failures. (This comprehensive ablation is now included in Appendix C.1, Table 7.)

Q4/W2: Can G-VR improve stronger AE architectures?

Reply: We observe consistent directional evidence across different architectures, but with varying magnitudes of gain. We maintain a conservative claim regarding the absolute performance boost.

Details:

  • Capacity-matched MRI AE (Parameter difference < 2%): SSIM improved from 0.93636 to 0.96610; Dice from 0.64225 to 0.69704.
  • Transformer AE (Single Seed, matched parameters):
Geometry InputSSIMRMSEDice
Zero-geometry0.9718190.0308850.801487
Correctly paired G-VR0.9725040.0304700.802132
Shuffled G-VR0.9715390.0309830.802033

Correctly paired G-VR outperforms both baselines, indicating the gain stems from geometric correspondences rather than extra capacity. However, the SSIM gain (+0.000685) is small. We present this strictly as potential evidence under the current setting and do not claim stable cross-seed improvements. (The revised Section 4.4 explicitly states this limitation.)

Q5/W4: Performance and computational cost on larger datasets, cross-datasets, and higher resolutions?

Reply: Results on large-scale and cross-datasets at 64³ are highly consistent; for 128³, introducing BFM significantly reduces time and improves accuracy.

Details:

  • Large-scale 64³: OASIS-1 (100 cases) SSIM 0.998076±0.001973; OASIS-2 (373 scans, 150 subjects) SSIM 0.998057±0.001876.
  • 128³ Comparison (10 cases):
OT MethodVertices/AxisTotal VerticesUpdatesAvg. SSIMAvg. RMSE
SDOT25716,974,593160.9660670.011929
BFM1937,189,057320.9920670.003719

(We have added this comparison as Table 2 in Section 4.2 of the revised manuscript.)

Q6/Q2: Can you add 3D generation experiments?

Reply: We did not add generation experiments in this rebuttal to strictly distinguish "generation" from "restoration/interpolation."

Details: G-VR represents the deformation structure of a given volume; latent interpolation performs geometric interpolation between two known deformations. Using G-VR as a geometric prior for generative models (e.g., diffusion models on the QC field) is a promising future direction but requires an independent evaluation protocol. We refrain from overclaiming in this rebuttal.


Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

Note to Reviewers and AC:

We have uploaded a Revised Manuscript (PDF) along with this rebuttal. All promised updates—including the newly introduced continuous optimal transport (BFM), the expanded downstream learning results in Section 4.4, and the comprehensive new Appendices C.1–G—are fully incorporated in the revised PDF. We kindly ask you to refer to it for detailed data and tables.

Core Clarification

Before addressing the specific questions, we would like to clarify the core positioning of G-VR.

Existing volumetric representations—such as voxel intensities, occupancy values, and implicit fields—are essentially scalar representations: they only record "how much mass is at each point," but fail to capture "how this mass is structurally organized by deformation." G-VR introduces the concept of geometric representation to volumetric data. By treating the volume as a deformation from a uniform reference domain to a target measure, we explicitly encode this structural deformation using a 3D tensor field.

This yields three key properties that scalar representations lack:

  1. Intrinsic Invertibility and Topology Preservation: The QC tensor is positive definite if and only if the mapping is orientation-preserving. "No folding" is a mathematical guarantee of the representation itself, rather than relying on post-hoc regularization. Scalar field interpolation lacks this constraint, making latent interpolations prone to topological artifacts.
  2. Structure-Preserving Interpolation: Positive definite tensors remain positive definite under barycentric interpolation (Eq. 4), ensuring the local metric remains valid after grid changes or resampling.
  3. Feature Learnability: All samples share the same reference mesh with vertex-wise correspondences, allowing the QC field to serve directly as geometric features for neural networks.

We appreciate the reviewers' questions, which have helped us further highlight these properties and introduce the Continuous Optimal Transport (BFM) for scalability. Below, we address each point in detail.


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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.