REVIEW 3 major objections 4 minor 5 references
3DOF+Quantization: 3DGS quantization for large scenes with limited Degrees of Freedom
T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash
Pith's one-line read The paper claims that quantizing 3D Gaussian positions in (1/ρ, θ, φ) makes projection error independent of distance and gains 5.3 dB over uniform x,y,z at 12 bits per coordinate.
desk verdict The radial 1/ρ quantization is a genuinely useful tweak, but the angular-quantization optimality is derived for a 360° projection, not the pinhole model 3DGS uses — and the single-scene, training-views-only evaluation doesn't close that gap. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the spherical-coordinate reparameterization t=1/ρ paired with the derivative calculation of the spherical projection p=P/||P||. The calculation shows ∂p/∂θ and ∂p/∂φ are O(1) while ∂p/∂ρ is O(1/ρ²), so inverting the radius converts a distance-dependent quantization error into a distance-independent one; the center/periphery split is the auxiliary mechanism that keeps the near field where the expansion's small-parameter ε breaks down.
What would settle it
Compute actual pinhole projection error for points at distances spanning both sides of R under uniform 1/ρ,θ,φ quantization and compare with uniform x,y,z; if far-field error is not roughly flat or if near-field error grows beyond the Cartesian baseline, the central bound is violated.
Extended reading notes
Core claim
The paper's central claim is a parameterization identity for 3DGS coordinate compression: under a 360-degree spherical projection, the derivative of the projected point with respect to the angles θ and φ is bounded (order one), while its derivative with respect to the radius ρ decays as 1/ρ². Substituting t=1/ρ makes the derivative with respect to t a bounded vector that does not depend on the point's position, so uniform scalar quantization of (1/ρ, θ, φ) is the natural near-optimal choice in the far field. The paper then handles the near field by splitting the scene: Gaussians within a central radius R, about 1.5 times the training-camera radius, are quantized uniformly in x,y,z, and only
Load-bearing premise
The proof assumes a 360-degree spherical projection and that every Gaussian is far from the camera cluster compared with the cluster's size (ε small); when that fails, the method resorts to a manually chosen center/periphery split radius, so the claimed distance-independence does not apply inside the center.
Editorial extensions
If this is right
- At 12 bits per coordinate, the spherical scheme yields 29.30 dB versus 23.96 dB for uniform Cartesian on the Garden training views.
- The error analysis gives a closed-form bound: after inverting radius, quantization of the three spherical parameters has projection error independent of distance, so peak bit savings come exactly where Cartesian uniform quantization wastes bits.
- Because the parameterization is applied after training, it can be layered onto codebook, structured, or entropy-coded 3DGS compression without retraining.
- The center/periphery split can be signaled cheaply: ordering center points first leaves only an index, or one bit per Gaussian (0.33 bit per coordinate), and still beats uniform quantization.
- The method is designed for 3DoF+ scenes but the conclusion anticipates adaptation to generic scenes where far-away content is viewed at low resolution.
Reading between the lines
- The same 1/ρ coordinate could be applied per-axis or per-level-of-detail in streaming, so bit rate can be made a smooth function of depth instead of a binary center/periphery threshold.
- A natural extension is to replace the tuned scalar split radius with a learned or closed-form threshold derived from the camera position covariance, eliminating the need to tune on the test scene.
- The argument suggests that any 3D representation with projection-based rendering could benefit from depth-adaptive coordinate quantization, not only Gaussian splats; the spherical-derivative calculation transfers to point clouds and meshes.
- For multi-cluster 6DoF capture, one could assign each Gaussian to its nearest camera cluster and apply the same inverse-radius parameterization per cluster; the paper's single-cluster analysis is the special case.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a coordinate quantization scheme for 3D Gaussian Splatting (3DGS) under a '3DoF+' assumption, in which camera positions are confined to a small region relative to the scene. Section 3 analyzes projection error, first for the pinhole model (Eq. 3) and then for a 360° spherical projection (Eqs. 5–6). The central theoretical claim is that radial position error produces projection error decaying as 1/ρ^2, so the radial coordinate ρ should be encoded as t=1/ρ and quantized uniformly, while θ and φ can also be quantized uniformly because their Jacobians are bounded. A center/periphery split with radius R separates the scene into uniformly quantized and spherically quantized zones. Section 4 reports PSNR on the Garden scene (Table 1: 29.30 vs 23.96 at 12 bits/coord) with ablations in Fig. 4.
Significance. The t=1/ρ transform is a clean and potentially useful idea, and the Appendix derivation is internally consistent for the spherical projection. The paper is also transparent about the center/periphery split being a separate mechanism. If the angular-quantization claim held for pinhole 3DGS, the method would be a simple, training-free complement to existing codecs. However, the spherical projection used in the derivation does not match the pinhole projection used by 3DGS, so the central theoretical claim for θ/φ is not established. The empirical support is limited to one scene, training views, and a radius R selected on the target scene.
major comments (3)
- [§3, Eq. (7) and Appendix A] The claim that uniform θ,φ quantization is optimal is derived for the 360° spherical projection p=fP/||P||, whose angular Jacobians are bounded by fρ/||P||≤f. 3DGS renders with the pinhole projection of Eq. (3). In the same spherical coordinates, the pinhole derivatives are ∂u/∂θ=f sec^2θ cosφ and ∂u/∂φ=-f tanθ sinφ, which are direction-dependent and unbounded as the ray approaches the image plane. Thus the O(1) bound in Eq. (7) does not transfer. The radial t=1/ρ transform does transfer (its effect scales as 1/ρ^2 in both models), but the theoretical justification for the angular part of the encoder is not established for 3DGS. Please re-derive for pinhole geometry, or provide a quantitative argument that the spherical model approximates pinhole error in the 3DoF+ setting.
- [§4, Table 1 and Fig. 4] The experimental section reports one scene (Garden), PSNR on training views, and a coordinate-only bit budget (no color, scale, SH, or split-mask bits). More importantly, the center/periphery radius R is tuned on the target scene ('bigger values of R did not improve...'), so the reported gain is not a fair comparison. Please provide a validation protocol in which R is fixed by theory or cross-validation, include at least a second scene, and report test-view PSNR and a full rate-distortion budget including the split information.
- [§3 'Center vs periphery' and §4] The split radius is defined as 'roughly twice' the camera distance in §3 but set to 1.5× in §4. Because the ablation (Fig. 4) shows that the center/periphery split is a major source of the improvement, this inconsistency matters. Report sensitivity of PSNR to R and give explicit guidance for choosing R without using the test set.
minor comments (4)
- [Abstract / §3] The abstract says projection error is proportional to squared inverse distance, but §3 shows transverse x,y errors scale as 1/z and only the radial z error scales as 1/z^2. Please qualify.
- [Eqs. (6)–(7)] In Eq. (6), p is the unit direction P/||P||, while Eq. (7) and the text refer to projection on a sphere of radius f. Normalize the definition and the derivative expressions consistently.
- [§4] PSNR is reported on training views; state this in the experimental setup and abstract to avoid overclaiming.
- [§3] Typo: 'but This gives' should be 'but this gives.' Also the center radius is 'roughly twice' in §3 but 1.5× in §4.
Circularity Check
No significant circularity; derivation is analytic and self-contained.
full rationale
The paper's central claim—that quantization of 1/ρ, θ, φ yields projection error roughly independent of distance—is derived analytically in Sec. 3 and Appendix A via Jacobians of a spherical projection. The derivation does not fit any parameter to the Garden scene: Eq. (8) follows from the chain rule and the far-field approximation ε = ||P0||/||P|| << 1. The proposed scheme is then compared against uniform x,y,z quantization as an external baseline. The center/periphery split radius R is set empirically (R = 1.5 × training-camera radius, with a note that larger R did not improve novel views); this is a tuned hyperparameter of the implementation, not a quantity predicted by the theory, so it does not make the prediction circular. There are no self-citations or imported uniqueness theorems. The main limitation—that the error analysis uses a 360° spherical projection while 3DGS uses pinhole projection—is a correctness/validity concern about transferring the bound, not a circularity of the derivation chain, and is explicitly acknowledged as an approximation by the authors. No circular step meets the evidentiary bar.
Assumptions & free parameters
free parameters (1)
- center/periphery radius R =
1.5 x radius of training camera positions
assumptions (3)
- standard math High-rate quantization: quantization of coordinates can be modeled as adding independent uniform noise to each coordinate.
- domain assumption Far-field approximation: Gaussians are far from the camera cluster compared to the cluster size (ε = ||P0||/||P|| << 1).
- ad hoc to paper The 360-degree spherical projection (points projected onto a sphere of radius f) is an acceptable error model for the pinhole projection used in 3DGS.
Cite this review
Pith. "Pith review of 3DOF+Quantization: 3DGS quantization for large scenes with limited Degrees of Freedom." pith.science (2026). https://pith.science/paper/IBXQO7DK
@misc{pith2026250906400,
author = {Pith},
title = {Pith review of: 3DOF+Quantization: 3DGS quantization for large scenes with limited Degrees of Freedom},
year = {2026},
howpublished = {\url{https://pith.science/paper/IBXQO7DK}},
note = {Machine review of arXiv:2509.06400}
}
read the original abstract
3D Gaussian Splatting (3DGS) is a major breakthrough in 3D scene reconstruction. With a number of views of a given object or scene, the algorithm trains a model composed of 3D gaussians, which enables the production of novel views from arbitrary points of view. This freedom of movement is referred to as 6DoF for 6 degrees of freedom: a view is produced for any position (3 degrees), orientation of camera (3 other degrees). On large scenes, though, the input views are acquired from a limited zone in space, and the reconstruction is valuable for novel views from the same zone, even if the scene itself is almost unlimited in size. We refer to this particular case as 3DoF+, meaning that the 3 degrees of freedom of camera position are limited to small offsets around the central position. Considering the problem of coordinate quantization, the impact of position error on the projection error in pixels is studied. It is shown that the projection error is proportional to the squared inverse distance of the point being projected. Consequently, a new quantization scheme based on spherical coordinates is proposed. Rate-distortion performance of the proposed method are illustrated on the well-known Garden scene.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[1]
T., Mildenhall, B., Verbin, D., Srinivasan, P
Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., and Hedman, P. (2022). Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 5470--5479
work page 2022
-
[2]
Chen, Y., Wu, Q., Cai, J., Harandi, M., and Lin, W. (2024). Hac: Hash-grid assisted context for 3d gaussian splatting compression. arXiv preprint arXiv:2403.14530
work page Pith review arXiv 2024
-
[3]
Kerbl, B., Kopanas, G., Leimk \"u hler, T., and Drettakis, G. (2023). 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. , 42(4):139--1
work page 2023
-
[4]
Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., and Dai, B. (2024). Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 20654--20664
work page 2024
-
[5]
Papantonakis, P., Kopanas, G., Kerbl, B., Lanvin, A., and Drettakis, G. (2024). Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques , 7(1):1--17
work page 2024
Reviewed August 4, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.