acceptodds
Withdrawn from ICLR 2027

Neural Explicit Shape Representation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing neural 3D representations separately address continuous surface modeling, efficient image synthesis, or object-specific compression. They do not jointly provide the explicit surface queries required by ray tracing and a structured output space that can be predicted across objects. We introduce NeuOctree, a neural explicit representation that directly predicts whether and where a ray intersects a surface, together with its normal and albedo. NeuOctree stores shared features at the corners of sparse octree voxels and interpolates them into a continuous local field. A lightweight decoder with recurrent intersection-guided sampling refines each query near the predicted surface, enabling direct surface rendering without mesh extraction or texture baking and accommodating open, non-watertight geometry. The same octree structure supports both compact per-object optimization and a shared octree CNN that constructs renderable assets from colored point clouds without per-object fitting. Across seven meshes and four storage budgets, NeuOctree achieves the lowest positional error in 24 of 28 comparisons while retaining competitive normal accuracy, and substantially improves joint geometry–appearance reconstruction over N-BVH and VQAD. The feedforward model constructs a representation in 0.05 seconds and transfers from Objaverse to ShapeNet without retraining. We further demonstrate ray-traced rendering and image-conditioned generation, showing that NeuOctree provides a common output space for neural inference and direct surface rendering.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Neural rendering, neural representation, mesh compression

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3GcVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NeuOctree, a neural surface representation that combines shared octree features with an MLP for direct ray-surface intersection, normal, and albedo prediction. Recurrent intersection-guided sampling improves query accuracy. The representation supports both per-object compression and feedforward construction, with additional demonstrations in ray tracing and image-conditioned 3D generation.

Critical strengths
  1. NeuOctree provides a compact neural representation that directly predicts ray-surface intersections and surface attributes, offering a useful interface for surface rendering.

  2. The method achieves strong geometric compression accuracy across multiple storage budgets. Its ray-tracing and 3D-generation demonstrations also illustrate the potential versatility of the representation.

Critical weaknesses
  1. Direct neural ray-intersection prediction has already been explored in N-BVH (SIGGRAPH 2024), while octree-based neural feature representations are well established (e.g., NGLOD, Octree-GS, and PlenOctrees). The proposed method largely combines existing ideas, and its distinctive technical contributions beyond N-BVH are not sufficiently established.

  2. For a representation-focused paper, the experiments provide surprisingly limited analysis of the representation itself. Key design choices, including shared corner features, octree structure, loss formulations, and the depth–accuracy–efficiency trade-off, lack systematic ablations. Without comparisons against simpler alternatives, such as sparse voxel features with an MLP, the necessity and advantages of the proposed design remain unclear.

Questions for authors

Q1. Section 3.4 describes construction from colored point clouds, but Appendix A.2 additionally uses mesh-rasterized occupancy and colors from eight rendered albedo views. Does test-time inference require access to the ground-truth mesh? Please clarify the exact input protocol and compare point-cloud-only, mesh-assisted, and full-input configurations.

Q2. The reported 50 ms excludes preprocessing and rendering. Please report the complete asset-construction time, including octree initialization, and provide actual ray-tracing latency or throughput.

Q3. Please evaluate shared versus alternative spatial encodings (e.g., sparse voxel grids), the contributions of key loss terms, and the trade-off among octree depth, accuracy, memory, and inference speed.

Q4. The ray-tracing and image-conditioned generation results are primarily qualitative. While SOTA performance is not necessarily required, quantitative evidence of rendering quality, efficiency, or structural advantages over alternative representations is needed to substantiate the claimed application benefits.

Q5. Ray-wise errors alone are insufficient to assess global geometric fidelity. Please reconstruct a point cloud from NeuOctree via multi-view ray queries and evaluate it against the original mesh using Chamfer Distance, F-score, and surface completeness.

Q6. Figures 10 and 11 do not clearly establish the advantages over the strongest baselines. Many results appear visually similar, and Figure 10 omits QNDF despite its strong quantitative performance. Please provide informative zoomed-in comparisons and geometric, normal, or albedo error maps.

Q7. The term "Neural Explicit Shape Representation" may be somewhat ambiguous, as the surface is represented through neural ray-intersection predictions rather than explicit geometry. Would a more specific title better reflect the paper's actual technical contribution?

Reviewer_LGc4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an octree-based neural 3D representation that predicts surface intersection points, normals, and albedo. It supports both per-mesh optimization and feed-forward prediction using a U-Net-style octree CNN, with a recurrent decoder iteratively refining intersection predictions. Good results are demonstrated on synthetic shapes. Its main limitation is its reliance on high-quality depth information to derive colored point clouds for constructing the octree feature structure. This limits its applicability to real-world RGB images, for which depth must be estimated and can be inaccurate.

Critical strengths
  1. The occupied-voxel feature field achieves good compression of shape meshes.

  2. The feed-forward architecture predicts the feature field across shape instances and datasets, enabling direct ray-traced rendering.

  3. The proposed method offers largely improved computational efficiency.

Critical weaknesses
  1. The method relies on occupied voxels initialized from input point clouds. How sensitive is the resulting neural representation to input sparsity or incompleteness or noisy depth? Does the image-conditioned generator help in this case, and how?

  2. The rendering results primarily show clean, isolated shapes. How well does the method extend to complex scenes, such as the Bonzai, City Block, and Andalusian Room used in N-BVH, particularly regarding intersection accuracy for small leaves and thin structures?

  3. Section 4.2 does not specify whether the input point clouds are obtained from mesh-rendered RGB-D images, presumably following Section 4.1. Would point clouds sampled directly from meshes (rather than RGB-D-derived) yield comparable performance using the feed-forward model?

  4. In Fig. 3, NGF appears to preserve finer details better, particularly around Einstein’s eyes, despite the proposed method achieving a lower quantitative error.

Questions for authors

How much does each recurrent refinement step improve intersection-point accuracy?

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

This paper presents NeuOctree, a neural explicit representation that models where a ray would intersect an object’s surface along with the surface normal and albedo information. Given an input point cloud of the object to encode, NeuOctree constructs an octree around it and assigns learnable features to voxel corners while optimizing a common decoder.

Critical strengths
  • The overall problem motivation is well-grounded; having a representation that can be directly renderable by a raytracing-based renderers is a big plus.
Critical weaknesses
  • Concern about quality of NeuOctree for rendering: One of the main selling points of NeuOctree is that it would be easy to render in traditional raytracing renders. However, as the authors point out on L.482-483, “the decoder exposes albedo rather than a full view-dependent material model.” If I understand correctly, this means that there may be little to no realistic view-dependent color changes in a renderer, which would greatly limit the applicability of NeuOctree. In Fig. 11, it also seems like qualitatively, NeuOctree sometimes has worse fine details compared to other methods such as VQAD.
  • The octree formulation / how the NeuOctree takes advantage of hierarchical organization is unclear. In Section 3.1 – 3.3, it seems like intersections are computed directly against leaf nodes of the octree structure, which can be inefficient. Additionally, one assumption the paper makes is that the recurrent intersection refinement will lead the model closer to the true surface intersection, but it is unclear to me if this must actually be the case; there doesn’t seem to be any pressure for each refinement to get monotonically closer to the true surface, or other guarantees of the likes.
Questions for authors
  • Can you show NeuOctree’s rendered outputs for an object with view-dependent appearance variation, such as a glossy object? Also, can you provide the rendering FPS for NeuOctree in comparison to other methods?
  • Can you clarify if there is provable guarantee that the recurrent intersection refinement will get closer to the true surface intersection over multiple calls?

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

Related papers

Discussion (0)

Sign in to comment.