acceptodds
Under review as a conference paper at ICLR 2027

P2Voxel: Pyramid Pivot Voxelization for 3D Mesh Tokenization

Abstract

Triangle meshes provide explicit and accurate surface geometry, yet their irregu- lar topology connectivity makes 3D mesh tokenization a geometric sampling problem: how to sample and organize geometric evidence into compact, structured and learnable tokens. Beyond field-centric volumetric sampling and edge-intersection surface sampling, we retarget mesh tokenization as local surface evidence sampling: representing each active voxel with compact local surface evidence that supports deterministic reconstruction. To this end, we introduce P2Voxel, a pyramid pivot voxelization framework for compact and reconstruction-aware mesh tokenization. P2Voxel is built on three key components. Under the Local Planarity assumption, Pivot Voxelization represents each active voxel with a surface pivot and an orientation sign, providing compact local evidence that can induce the corner values required for deterministic reconstruction. Under the Spatial Complexity assumption, Pyramid Pivot Voxelization exploits the spatial non-uniformity of real surfaces by allocating finer pivot tokens to geometrically complex regions while keeping smooth regions coarse and compact. Under the Block Reconstructability ity assumption, a Pyramid VAE learns compact multi-resolution latent codes over locally reconstructable pivot blocks, avoiding the need to model the entire high-resolution voxelized shape as a dense global field. Together, these designs convert meshes into compact, structured, and learnable pyramid pivot tokens, enabling efficient mesh reconstruction for downstream 3D tasks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
3D mesh tokenization, voxelization, surface reconstruction, 3D representation learning

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 4 reviews, above 85% of the others.
Reviewer_SZWCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes P2Voxel, a mesh tokenization framework combining a compact pivot and orientation sign representation, geometry adaptive pyramid sampling, and a block-wise Pyramid VAE. Pivot tokens induce local scalar values for surface extraction, while adaptive resolution allocation reduces the number of sampling primitives. Experiments evaluate direct reconstruction, VAE reconstruction, cross-dataset generalization, and masked latent completion.

Critical strengths
  1. The framework has a clear motivation and connects representation design, adaptive sampling, and learned compression within a coherent pipeline.
  2. The reported results demonstrate promising reconstruction and complexity trade-offs across ABO, Objaverse, and Wild. The VAE experiments additionally provide evidence of generalization beyond the training dataset.
  3. The paper includes useful implementation details, resolution and allocation ablations, and a masked latent completion experiment.
Critical weaknesses
  1. Moving a projected pivot back inside its voxel may change the recovered surface normal. The paper does not quantify how often this occurs or how much reconstruction error it introduces.
  2. The experiments combine pivot encoding, adaptive sampling, and VAE compression, making their individual benefits difficult to identify. The implementation and training settings of the VAE baselines also need clarification.
  3. Token counts and per-token dimensions do not capture the full storage cost. The comparisons lack total latent sizes and the overhead of spatial information and the sampling map.
  4. The method relies on cross-resolution mesh stitching and hole repair, but their effects on mesh quality and runtime are insufficiently evaluated.
  5. The “Explicit 3D Representations for Generation” section in Related Works omits recent flow and diffusion-based mesh generation methods, such as MeshFlow, LATO, LATO.2, and Nexus. Discussing these methods would help clarify how P2Voxel differs in its representation design and handling of mesh connectivity.
Questions for authors
  1. How are pivots outside their voxels handled, and what reconstruction errors does this introduce?
  2. Can the authors clarify the baseline settings and compare different local encodings using the same adaptive sampling layout?
  3. Can the authors report total encoded and latent sizes and compare reconstruction quality at matched storage budgets?
  4. Can the authors report mesh defects before and after stitching and repair, together with the runtime of the current complete pipeline?
Reviewer_m6pzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

P2Voxel represents each active voxel using a surface pivot and orientation sign, from which a local plane induces corner distances for Marching Cubes. Adaptive pyramid allocation concentrates tokens in complex regions, and a blockwise VAE compresses the representation. Reconstruction, cross-dataset evaluation, and masked latent completion demonstrate favorable compactness and fidelity.

Critical strengths
  • The compact, plane-inspired descriptor connects local surface evidence directly to corner-distance reconstruction, avoiding independent storage of many field or edge constraints when its geometric conditions hold.
  • Adaptive allocation and blockwise compression produce a useful fidelity–primitive-count trade-off, with cross-dataset reconstruction and masked completion supporting a representation contribution in its own right.
Critical weaknesses
  • Feasibility repair introduces an additional approximation. Equation 3 obtains orientation from the center-to-pivot vector, requiring an orthogonal projection for exact planar recovery. Even a planar patch can have that projection outside its voxel; Appendix A's clamping then changes the recovered normal. The appendix acknowledges center-coincident ill-conditioning too, but leaves its handling unclear. This does not invalidate approximate reconstruction, but local planarity alone does not establish the claimed recovery conditions.
  • Primitive savings do not fully establish learned-code savings. Table 2 reports VAE fidelity without a comparable total per-shape latent budget. Tables 8/9 assign 1024 latent scalars per encoded block, plus allocation metadata. Reporting the complete budget would clarify how geometric sample reductions translate into compression.
  • While the absence of a full generative model is not itself a weakness, it renders the contributions weaker and less convincing.
Questions for authors
  • How often is an orthogonal pivot infeasible, what error does its repair introduce, and how are center-coincident or near-center pivots handled?
  • What are the full per-shape latent/storage budgets and post-stitch topology-change or interface-failure rates?
Reviewer_trwGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes P2Voxel, a compact mesh tokenization framework based on local surface evidence. The method represents each active voxel using a surface pivot point and an orientation sign, from which corner signed-distance values can be reconstructed for Sparse Marching Cubes. To improve efficiency, Pyramid Pivot Voxelization adaptively allocates different spatial resolutions according to local geometric complexity. The authors further introduce a block-wise Pyramid VAE to compress the resulting multi-resolution representation into learnable latent codes.

Experiments on ABO, Objaverse, and in-the-wild meshes demonstrate favorable reconstruction accuracy and token efficiency compared with SDF-based and dual-contouring representations. Additional experiments investigate VAE reconstruction quality, sampling-resolution trade-offs, and downstream masked latent completion.

Critical strengths
  • Compact and effective geometric representation. Representing local surface evidence using a pivot point and an orientation sign is a simple and interesting alternative to conventional SDF and dual-contouring representations. Combined with adaptive pyramid sampling, the method achieves a favorable trade-off between reconstruction fidelity and the number of geometric primitives. The improvements are consistent across multiple datasets.

  • Comprehensive evaluation of reconstruction and representation learning. The paper evaluates both deterministic reconstruction and VAE-compressed reconstruction, demonstrating that the proposed representation retains competitive geometric accuracy after compression. The cross-dataset experiments and masked latent completion results provide additional evidence that the representation is suitable for downstream learning.

Critical weaknesses
  • Insufficient comparison of actual representation efficiency. The primary efficiency metric is the number of sampled geometric primitives. However, this does not fully capture the storage and computational requirements of the proposed representation, which additionally requires spatial coordinates, a multi-resolution sampling map, and cross-resolution reconstruction. Moreover, the highest-quality Pyramid-1024 configuration uses a higher maximum spatial resolution than several compared configurations. Comparisons at matched memory budgets, latent sizes, and reconstruction resolutions would better establish the claimed efficiency advantages.

  • Limited analysis of geometric validity and generality. The proposed pivot representation relies on local planarity, which may not hold near thin structures, sharp features, or multiple nearby surface sheets. Although the authors investigate resolution-dependent artifacts, the pyramid reconstruction additionally requires heuristic cross-resolution stitching and boundary repair. The paper does not sufficiently quantify how reliably these operations preserve topology and watertightness. Furthermore, the evaluation relies on meshes converted into watertight surfaces, leaving the applicability to more general mesh inputs unclear.

Questions for authors

Can the authors provide comparisons at matched storage or latent-token budgets, accounting for the pyramid sampling map and spatial indexing overhead? End-to-end runtime comparisons against existing tokenization approaches would also help clarify the practical efficiency improvements.

How frequently does cross-resolution stitching introduce topological defects, non-manifold edges, or disconnected components? Can the authors provide quantitative topology-preservation metrics for the final pyramid reconstruction, particularly on objects containing thin or closely spaced surfaces?

Can the authors clarify the contribution of the Pyramid VAE compared with alternative latent compression approaches at the same compression ratio? The current masked latent completion experiment primarily compares learned prediction against zero filling, which does not fully establish an advantage over other learnable representations.

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

This paper proposes P2Voxel, a representation for 3D meshes based on a surface pivot and an orientation sign in each active voxel. The pivot induces a local plane from which corner values can be computed for sparse Marching Cubes reconstruction. A complexity-guided pyramid allocates finer voxels to selected regions, and a block-wise Pyramid VAE encodes the resulting tokens. The evaluation covers direct mesh reconstruction and reconstruction after VAE compression on ABO, Objaverse, and Wild, together with masked latent completion on ABO.

Critical strengths
  1. Strong reconstruction results with fewer sampling primitives. In Table 1, Pyramid-1024 achieves the best reported Chamfer distances, EMD, and F-score on all three datasets while using fewer primitives than Pivot-512. This directly supports the value of adaptive allocation and is the main reason I lean toward acceptance.

  2. A compact representation with a clear geometric interpretation. The pivot-and-sign construction gives a concrete way to derive corner values from local surface evidence. The Pyramid VAE results and masked latent completion experiment further suggest that the representation can support learned compression and prediction, beyond direct deterministic reconstruction.

Overall assessment (Weak Acceptance): I lean toward acceptance. The paper presents a geometrically interpretable representation and reports consistently strong reconstruction results with fewer sampling primitives across three datasets. The unresolved issues limit how strongly I can endorse the storage-efficiency and final-mesh robustness claims, but they appear amenable to clarification and targeted additional measurements. The authors’ answers to the questions above, particularly the Table 1/Table 4 discrepancy and complete representation cost, could change my assessment.

Critical weaknesses
  1. The full token-efficiency claim is not yet measured. The main comparison reports primitive counts and per-primitive dimensions, but the complete serialized representation cost, including any explicitly stored spatial or block-layout metadata and the sampling map, is not reported. Voxel indices recoverable from pivot coordinates should not be counted again. The VAE architecture specifies the latent size per block, yet the paper does not report the total per-shape latent budget or a matched-bitrate comparison. Thus, the reported results establish an advantage in primitive count more clearly than an advantage in complete storage cost.

  2. Important reconstruction and reporting details remain unresolved. Cross-resolution reconstruction relies on stitching and repair, but the paper does not quantify final-mesh validity, such as open boundaries, non-manifold edges, or self-intersections. In addition, Table 1 and Table 4 give different ABO reconstruction metrics for Pyramid-512/S2 despite reporting the same allocation ratios and 135,736 primitives. This may have an experimental explanation, but the current text does not provide one.

Questions for authors

Clarification requests that could affect my assessment:

  1. Table 1 reports ABO Pyramid-512 with 135,736 primitives and allocation ratios [0.85, 0.10, 0.05]. Table 4 reports S2 with the same count and ratios but different reconstruction metrics. Were the evaluation sets, preprocessing, or other settings different? Please explain the discrepancy.
  2. What is the complete storage cost per shape, including pivot values, any explicitly stored spatial or block-layout metadata, and the sampling map, without double-counting information recoverable from pivot coordinates? For the VAE results, how many latent scalars or bytes are used per shape, and how does quality compare at matched total bitrates?
  3. After cross-resolution stitching and repair, what proportion of output meshes have open boundaries, non-manifold edges, or self-intersections? How often does the repair stage alter topology?
  4. Does Table 11 measure the pipeline used for the main results? If so, where are stitching and repair costs included, and why is resolution lifting listed?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.