acceptodds
Withdrawn from ICLR 2027

Is Gaussian Enough? On the Local Specialization of Kernels in 3D Splatting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Recent 3D splatting methods have extended Gaussian primitives toward increasingly expressive and adaptive kernel formulations. However, it remains unclear whether greater within-family adaptability makes kernel families effectively interchangeable or whether they retain distinct local representation biases. Existing studies primarily rely on scene-level reconstruction metrics, potentially obscuring heterogeneous behavior within a scene. We systematically characterize local cross-family specialization across Gaussian, generalized exponential, and deformable radial kernels using real-scene analysis, structure-conditioned characterization, controlled synthetic manipulations, and robustness controls. We find substantial local complementarity: different families achieve lower reconstruction error in different regions, with relative performance varying systematically with local scene structure. Controlled structural manipulations produce reproducible changes in cross-family performance, while specialization persists across spatial scales and under a matched primitive budget. These results show that within-family adaptability does not necessarily eliminate cross-family representation differences, motivating representations that account for compatibility between kernel families and local scene structure.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Neural Scene Representation, Novel View Synthesis, Adaptive Kernels, Kernel Specialization, Representation Bias, Computer Vision

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

This paper investigates local reconstruction differences among three splatting kernel families: 3DGS, GES, and DRK. Through patch-level analysis, controlled synthetic experiments, and robustness control tests, the authors show that different kernel families perform better in different local regions, and these differences are related to scene structure.

Critical strengths

The paper provides a systematic local analysis of existing 3D splatting representations, going beyond conventional scene-level reconstruction metrics.

The controlled experiments and robustness analysis provide useful evidence of local differences between kernel families.

The paper is easy to follow.

Critical weaknesses

My main concern is the limited novelty and contribution. The paper mainly analyzes existing kernel representations using standard evaluation and statistical tools. Though the observations are interesting, the conclusions are relatively intuitive, and the work offers limited new methodological insights.

Limited experimental coverage. The real-world evaluation includes only three scenes. It is unclear whether the reported patterns generalize to more diverse scenes and reconstruction settings. The qualitative comparisons are also limited, making the practical significance of local specialization difficult to assess.

Questions for authors

Can the authors extend the evaluation to more diverse real-world scenes to verify the generality?

Can the authors provide continuous novel-view rendering demonstrations to show whether the local advantages remain consistent across viewpoints?

Reviewer_m9hUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates an interesting and relevant question: Do increasingly expressive splatting kernels become interchangeable, or do different kernel families retain systematic advantages in different regions of a 3D scene? The authors compare three representations: 3D Gaussian Splatting (3DGS), Generalized Exponential Splatting (GES), and Deformable Radial Kernel (DRK) Splatting, using patch-level reconstruction errors, structural image descriptors, controlled synthetic scenes, and primitive-budget controls. The central empirical observation is compelling: even when 3DGS performs best globally, GES and DRK can achieve lower reconstruction error in substantial portions of the same scene.

Critical strengths

1. Well-motivated research question. Most splatting benchmarks emphasize scene-level PSNR, SSIM, and LPIPS. Investigating spatially varying relative performance addresses a meaningful limitation of global evaluations.

2. Clear experimental organization. The four research questions map naturally to the evaluation sections. The paper separates descriptive real-scene analyses, controlled synthetic experiments, and robustness checks.

3. Careful acknowledgment of limitations. The authors explicitly recognize that their comparisons do not causally isolate the kernel functional form and that the oracle is diagnostic rather than realizable. This restraint improves the credibility of the paper's stated empirical conclusions.

Critical weaknesses

W1. The paper demonstrates that different independently trained representations win in different image patches. However, it does not establish whether these differences are specific to kernel families or are a more general consequence of training different reconstruction models.

For example, suppose three independently trained 3DGS models were compared using exactly the same local winner and oracle procedure. They might also exhibit substantial variation in their locally preferred regions. Also, the theoretical discussion in Eq. (3) considers the minimum achievable approximation error for each family. The experiments instead measure the errors attained by particular training runs. These are not equivalent.


W2. Differences in capacity and optimization remain major confounding factors. The methods are trained with different optimization schedules, densification parameters, and final primitive counts. The primitive-count imbalance is substantial. GES contains the Gaussian kernel as a special case. Under genuinely comparable representation spaces and equal budgets, a strictly more expressive family should not have worse optimal approximation error than a nested, less expressive family. The paper correctly avoids claiming direct causal isolation of kernel shape. Nevertheless, its interpretation of persistent representation biases should discuss this nesting argument more explicitly.

W3. Structural correlations do not necessarily establish meaningful structure-dependent preference The structure-conditioned analysis is a central contribution, but I have three concerns. First, the reported Spearman correlations are often modest. For example, GES–3DGS correlations range from approximately 0.069 to 0.193 across the six descriptors and three scenes. DRK–3DGS correlations are stronger but still moderate.

Questions for authors

Please refer to the weaknesses

Reviewer_qrk3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose that newer splatting kernels are more flexible than the Gaussian kernel. GES learns how fast each kernel fades out, and DRK learns a deformable footprint. The paper asks whether that flexibility makes the kernel families interchangeable, or whether each one still does better on certain kinds of local structure. In their results, 3DGS is best scene-wide on all three Mip-NeRF 360 scenes. Even so, GES or DRK wins in some local patches. Picking the best family per patch would cut error relative to 3DGS.

Critical strengths
  • The authors show an observation that even though 3DGS gives best overall fidelity for a scene, it does not necessarily gives the best fidelity in all areas, which is already an established observation with many previous works on 3D reconstruction and novel view synthesis. The authors visualized the areas where each kernel can give the best result, but did not formalize this observation into any statistically relevant conclusion.
Critical weaknesses
  • The empirical evidence provided by the authors is not strong enough. Per-patch training result can depend heavily just on the seed and initialization selected which the authors did not account for as far as I know. It is hard to draw any reliable conclusion without accounting for the potential variations produced by the stochastic training process.
  • The structure correlations may just reflect scale, not specialization. Every method makes larger errors in detailed patches. If DRK is worse overall (it has the lowest PSNR on every real scene), then DRK's error minus 3DGS's error will grow with detail even if DRK has no structure-specific weakness.
Questions for authors
  • The authors say themselves that this shows association, not causation. The three methods also differ in parameterization and training, so the effects can't be attributed to the kernel shape alone. Do the authors care to elaborate more on how they plan to solve these issues?
  • The evidence is limited: only three real scenes, and the structure measures are 2D image statistics standing in for 3D structure. Can the authors provide stronger empirical results?

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

Related papers

Discussion (0)

Sign in to comment.