{"id":"7e8cd23c-03c2-48e8-8a2d-6e31625f2a6e","arxiv_id":"2509.06400","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Under the 3DoF+ scenario (cameras confined to a small region), quantizing Gaussian positions with uniform steps on 1/ρ gives better rate-distortion than uniform x,y,z quantization.","lead":"This paper proposes a way to compress 3D Gaussian Splatting models by storing Gaussian positions in spherical coordinates and quantizing the inverse distance. If valid, it offers a cheap bitrate win for large scenes viewed from a small camera region, common in VR and telepresence.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The angular-quantization optimality is derived for a 360° spherical projection, not for the pinhole projection 3DGS actually uses; the O(1) Jacobian bound in Eq. (7) does not transfer.","rationale":"The reader identifies the switch from pinhole to spherical projection and the unverified far-field assumption as the weakest point; my stress-test sharpens this into the main load-bearing concern. The appendix derivation is internally consistent for the spherical model, and the 1/ρ radial component is also consistent with pinhole parallax (error ∝ 1/ρ^2). However, the angular component relies on the spherical-projection Jacobian being O(1) in all directions, which is false under pinhole projection with arbitrary camera rotations: the pinhole angular Jacobian contains sec^2 of the off-axis angle and varies with camera orientation. This gap means the theoretical optimality of uniform θ/φ quantization is unproven for the actual renderer. The empirical evidence does not close the gap because the 'w/o spherical' ablation does not isolate angular quantization from the radial/center-periphery effect, and R is tuned on the test scene. Since the reader already marked the paper CONDITIONAL on essentially this weakness, my read does not change the verdict; a targeted pinhole-aware comparison would settle whether the mismatch is practically important.","tokens_in":5099,"tokens_out":10116,"duration_ms":121324,"concrete_test":"Implement a pinhole-aware variant of the quantizer: keep the same bit budget, the t = 1/ρ quantization, and the center/periphery split exactly as proposed, but replace uniform θ/φ quantization with nonuniform steps inversely proportional to the RMS pinhole Jacobian ‖∂(u,v)/∂(θ,φ)‖, averaged over the Garden training cameras using their actual rotations. Compare PSNR at 12 bits/coord against the proposed uniform-θ/φ scheme. If the pinhole-optimized angular quantizer does not improve over uniform, the spherical-vs-pinhole mismatch is empirically harmless; if it does improve, the paper's theoretical justification for the angular component is incorrect and the reported gain is not attributable to the claimed distance-independence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central argument for uniform θ/φ quantization rests on the spherical projection p = P/||P|| (Eqs. 6–7), whose angular Jacobians are bounded by ρ/||P|| ≤ 1. 3DGS, however, renders with pinhole projection u = f x/z, v = f y/z, after rotating each camera into its own frame. For a camera aligned with the world z-axis, the pinhole Jacobian of u with respect to θ is ∂u/∂θ = f cosφ sec^2θ (and similarly for φ); this is not the direction-independent, O(1) error of Eq. (7). For a fixed quantization step Δθ, the pixel error grows as the square of the secant of the angle between the Gaussian direction and the camera optical axis—e.g., a factor of 2 at 45° off-axis and 4 at 60°—and it also depends on the camera orientation, which the local-referential derivation never introduces. The radial part (t = 1/ρ) does transfer to pinhole geometry: parallax error scales as 1/ρ^2, so quantizing t yields distance-independent error. But the angular part of the claimed optimality is for a different camera model. Consequently, the paper has not established that uniform θ/φ quantization is near-optimal for 3DGS; the observed gain in Table 1 could stem mostly from the radial reallocation and the center/periphery split, whose radius R is tuned on the test scene, rather than from the angular coordinate scheme. This is a real correctness risk for the central claim, even though it does not prove the method fails empirically.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":5511,"tokens_out":10446,"duration_ms":106366,"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":[{"comment":"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.","section":"§3, Eq. (7) and Appendix A"},{"comment":"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.","section":"§4, Table 1 and Fig. 4"},{"comment":"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.","section":"§3 'Center vs periphery' and §4"}],"minor_comments":[{"comment":"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.","section":"Abstract / §3"},{"comment":"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.","section":"Eqs. (6)–(7)"},{"comment":"PSNR is reported on training views; state this in the experimental setup and abstract to avoid overclaiming.","section":"§4"},{"comment":"Typo: 'but This gives' should be 'but this gives.' Also the center radius is 'roughly twice' in §3 but 1.5× in §4.","section":"§3"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline case. The radial-coordinate idea is solid, but the angular claim's camera-model mismatch is central. If the authors cannot provide a pinhole derivation, they should reframe the method as a heuristic and strengthen the experiments. I would not accept in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has one solid idea and one unproven half. The solid idea is radial reparameterization: under the 3DoF+ assumption, uniform quantization of 1/ρ gives a projection error that does not grow with distance, and the derivation in Appendix A is internally consistent. The unproven half is the angular claim. The paper derives uniform θ,φ quantization from a 360° spherical projection, but 3DGS renders with pinhole projection. For pinhole, the pixel-space Jacobian in θ and φ is not bounded: off-axis, a fixed angular step produces pixel errors that grow with the square of the secant of the angle, and the effect depends on optical-axis orientation. The stress-test note is correct. That means the central \"spherical coordinates\" optimality argument does not transfer as written.\n\nWhat is new: the specific 1/ρ plus center/periphery split for 3DGS positions in tight-baseline scenes. The 1/ρ trick itself is old in depth/point-cloud coding, and the paper does not cite that line of work, but the context and the projection-error derivation are new here. The ablation is a good-faith attempt to separate the two ingredients, and the paper is short and readable.\n\nSoft spots, in order of damage:\n\n1. The abstract says projection error is proportional to squared inverse distance, full stop; Section 3 says x,y scale as 1/z and z as 1/z². The abstract overreaches.\n\n2. The evaluation is one scene, training views only, and only position bits. No held-out views, no full-model bitrate, no overhead for the center/periphery flag. At 12 bits/coord the reported gap is large (29.30 vs 23.96), but the angular part is not isolated from the radial reallocation, so the table does not yet support the paper's central optimality claim.\n\n3. R, the center/periphery radius, is set to 1.5× camera radius and the footnote says bigger values did not improve novel views; that is a fitted parameter, not a prediction.\n\nNone of this means the method fails. It could still be useful in practice. But as written, the paper has established the radial part and not the angular part. A revision should either derive the pinhole Jacobian for a distribution of camera orientations, or present a fuller empirical case with several scenes, held-out views, overhead included, and an ablation that isolates the angular quantization gain.\n\nFor peer review: send it. The flaw is addressable, the problem is real, and the paper is honest enough to admit the split is needed. It just needs sharper claims and better experiments.","headline":"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.","tokens_in":5972,"tokens_out":2967,"would_cite":true,"duration_ms":34249,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["3D Gaussian Splatting","quantization","spherical coordinates","rate-distortion","3DoF","novel view synthesis","point cloud compression","position coding"],"falsifier":"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.","tokens_in":5012,"feed_emoji":"📐","tokens_out":6909,"duration_ms":66872,"temperature":0.7,"pith_summary":"Most 3D scene compression quantizes Gaussian positions with uniform step sizes in x,y,z, spending the same number of bits on a nearby point and a distant one. This paper argues that in a 3DoF+ setting—cameras confined to a small zone in front of a large scene—the projection error caused by Cartesian quantization grows with the squared inverse distance to the camera cluster. The proposed fix is to reparameterize positions as spherical coordinates and uniformly quantize 1/ρ, θ, and φ instead, which makes the dominant projection error roughly constant for far-field points. Because that far-field assumption does not hold near the cameras, the method keeps a small central region on uniform Cartesian quantization and applies the spherical parameterization only in the periphery. On the Garden scene it raises PSNR from 23.96 to 29.30 dB at 12 bits per coordinate, matching or exceeding uniform Cartesian at all tested rates, while remaining a post-training operation that can be combined with other 3DGS compression schemes.","feed_headline":"Quantize by 1/distance, not x,y,z: 5 dB gain on 3D scenes","feed_subtitle":"In a large scene captured from one zone, uniform quantization of inverse radius and angles stops far-away Gaussians from wasting bits.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Defines the 3D Gaussian Splatting model and rendering pipeline whose position coordinates are being quantized.","marker":"[Kerbl et al., 2023]"},{"why":"Supplies the Garden scene and its training-view setup used for the rate-distortion measurements.","marker":"[Barron et al., 2022]"},{"why":"Establishes coordinate quantization as a memory-reduction strategy for 3DGS, the context the proposed bit-allocation scheme extends.","marker":"[Papantonakis et al., 2024]"}],"fun_headline_variants":["Quantize 1/radius, not xyz: better 3DGS compression","Spherical coords cut 3DGS bitrate for large scenes","3DGS: near-optimal quantization via inverse distance","Why 3DGS codecs should use 1/ρ, θ, φ quantization","5 dB gain in 3DGS from spherical quantization scheme"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Quantize 1/radius, not xyz: better 3DGS compression","Spherical coords cut 3DGS bitrate for large scenes","3DGS: near-optimal quantization via inverse distance","Why 3DGS codecs should use 1/ρ, θ, φ quantization","5 dB gain in 3DGS from spherical quantization scheme"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00095,"raw_usage":{"total_tokens":3910,"prompt_tokens":783,"completion_tokens":3127,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":3029}},"tokens_in":527,"tokens_out":3127,"duration_ms":27686,"temperature":1.0,"reasoning_tokens":3029,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:38:36.502179+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the 3D Gaussian Splatting model and rendering pipeline whose position coordinates are being quantized."},{"cited_title":"T., Mildenhall, B., Verbin, D., Srinivasan, P","cited_arxiv_id":null,"evidence_quote":"Supplies the Garden scene and its training-view setup used for the rate-distortion measurements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes coordinate quantization as a memory-reduction strategy for 3DGS, the context the proposed bit-allocation scheme extends."}],"review_version":1}