{"id":"99d9df4a-d1c9-4239-8cdc-99fd17103be7","arxiv_id":"2412.11752","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new 2D planar kernel primitive with learnable radial bases, mixed L1/L2 norms, and edge sharpening generalizes Gaussian splatting and claims better rendering quality with fewer primitives.","lead":"The paper introduces Deformable Radial Kernel (DRK) splatting, a new primitive for 3D scene rendering that replaces round Gaussian blobs with flexible, star-shaped kernels able to model sharp edges and corners. It reports sharper images with fewer primitives than Gaussian splatting on a new test set, but its advantage shrinks or reverses on standard benchmarks when camera poses are imperfect.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unqualified SOTA claim is undermined by the paper's own robustness data: under 1e-3 pose noise DRK's PSNR drops ~6 dB and falls below 3D-GS, so the advantage is conditional on near-exact camera poses.","rationale":"The reader's conditional verdict is well supported. The mathematical construction is a legitimate extension: Eq. (6) reduces to the 2D Gaussian for K=4 with orthogonal bases, and the L1/L2 blending and sharpening functions are differentiable and monotone, so the core primitive is sound. The DiverseScenes results are internally consistent and show a large margin over baselines. However, the strongest empirical claim is not robust across the paper's own experiments. Table S6 is decisive: at 1e-3 pose noise, DRK loses about 6 dB while 3D-GS loses about 1 dB, flipping a +3.2 dB lead into a -1.9 dB deficit. Table S5 shows the same pattern on Tanks & Temples, a standard real-world benchmark with imperfect COLMAP poses. This is not a disagreement with current consensus; it is an internal inconsistency between the abstract's unconditional language ('state-of-the-art', 'dramatically reducing primitive count') and the supplementary's own limitation analysis. I also note the paper states that the derivation of Eq. (11) is in the supplement, but the supplement contains no such derivation, and Sec. 4.1's definition of the polar angle via arccos discards the sign of v and is at least a notational ambiguity that should be corrected to atan2. Neither issue changes the conditional verdict, since the core method is promising and the pose-sensitivity concern is already acknowledged in the supplement. The absence of released code/data and of error bars further prevents unconditional acceptance, but it is not itself a mathematical objection.","tokens_in":18980,"tokens_out":17243,"duration_ms":169029,"concrete_test":"Re-run the Table S6 protocol on DiverseScenes with pose noise std 1e-3, but optimize camera poses jointly during training for both DRK and 3D-GS using identical pose-refinement settings. If DRK's PSNR still trails 3D-GS by about 2 dB, the fragility is structural and the abstract's SOTA claim should be narrowed; if DRK recovers and surpasses 3D-GS, the concern is specific to fixed noisy poses and the claim should carry an explicit pose-accuracy qualifier.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract) is that DRK achieves state-of-the-art rendering quality while dramatically reducing primitive count. For this to hold across real captures, the advantage must not evaporate under small perturbations of the estimated camera poses. The paper's own Table S6 shows the opposite: on DiverseScenes, adding pose noise with std 1e-3 makes DRK's average PSNR fall from 37.58 to 31.59, while 3D-GS only falls from 34.41 to 33.44; DRK therefore goes from +3.2 dB ahead to -1.9 dB behind. At 2.5e-3 and 5e-3 the gap widens further. Table S5 corroborates this on Tanks & Temples, where COLMAP poses are less reliable: DRK reaches 20.41 PSNR versus 21.09 for 3D-GS, with the authors attributing the drop to camera error. Because COLMAP pose errors in natural captures often exceed 1e-3, the unqualified 'outperforms existing methods' claim is not supported in a common real-world regime. The method is a credible primitive, but the headline result is conditioned on near-perfect poses and must be stated as such.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Deformable Radial Kernel (DRK) splatting, a planar primitive that generalizes 2D Gaussian splatting by using K learnable radial bases, an L1/L2 norm blend controlled by η, and a piecewise-linear sharpening parameter τ. A rendering pipeline is proposed with ray-plane intersection, polygon-based tile culling, a low-pass filter, and a cache-sorting scheme. Experiments compare DRK against 3D-GS, 2D-GS, 3D-HGS, and GES on a new DiverseScenes benchmark, Mip-NeRF360, NeRF-Synthetic, and Tanks & Temples, and the paper also demonstrates mesh-to-DRK conversion. The abstract claims state-of-the-art rendering quality with dramatically reduced primitive counts.","tokens_in":19342,"tokens_out":7156,"duration_ms":58962,"significance":"If the claims are properly scoped, DRK is a useful extension of Gaussian splatting: the derivation that a 2D Gaussian is a special case (Sec. 4.1) is sound, the L1/L2 blending and sharpening give the kernel additional expressive power, and the rasterization contributions (polygon culling, cache-sorting, low-pass filtering) are nontrivial. The authors also contribute a new benchmark. However, the headline SOTA claim is not supported on Mip-NeRF360 and is fragile under pose noise, so the significance depends on whether the method is presented as a specialized representation for accurate-pose settings.","major_comments":[{"comment":"The abstract's claim that DRK 'achieves state-of-the-art performance while dramatically reducing primitive count' is not supported on the standard Mip-NeRF360 benchmark: Table 3 shows DRK's PSNR (26.76) is lower than 3D-GS (L) (26.94), and its primitive count (952K) is higher than 3D-GS (811K). The advantage there is primarily in LPIPS/SSIM, and the dramatic reduction only holds for the S2/S1 variants on DiverseScenes. Please revise the claims to state the specific regimes where DRK is superior.","section":"Abstract; Sec. 5.2, Tables 1 and 3"},{"comment":"The robustness evaluation shows that DRK's advantage vanishes under realistic pose error: on DiverseScenes, adding Gaussian pose noise with std 1e-3 lowers DRK's average PSNR from 37.58 to 31.59, while 3D-GS only drops from 34.41 to 33.44 (Table S6); consistent with this, DRK underperforms 3D-GS on Tanks & Temples (20.41 vs 21.09, Table S5). Because COLMAP pose errors in natural captures often exceed 1e-3, the unqualified 'outperforms existing methods' claim in the abstract is not justified. The paper should either incorporate pose uncertainty into the representation or clearly state that the method targets settings with near-exact camera poses.","section":"Sec. S3, Tables S5 and S6"},{"comment":"The claim of 'dramatically reducing primitive count' is not consistent across experiments: full DRK uses 260K primitives on DiverseScenes (vs 346K for 3D-GS) and 952K on Mip-NeRF360 (vs 811K for 3D-GS), so the reduction is modest or reversed. The sparse variants (S1, S2) do reduce counts, but with lower PSNR. We recommend reporting results at matched primitive counts to substantiate the efficiency claim.","section":"Sec. 5.2, Tables 2 and 3"},{"comment":"The calibrated radial length is defined as sc_k = s_k sqrt(-log(Ψ^{-1}(e^{-32}/o))). The quantity e^{-32} appears to be a typo: if it is meant to encode the 3-σ boundary, the threshold should be exp(-9/2) or exp(-9) rather than exp(-32). Please clarify the derivation in the supplementary and ensure the culling radius is consistent with the kernel's actual support; an incorrect culling radius can affect both correctness and the claimed speedups.","section":"Sec. 4.4, Eq. (11)"}],"minor_comments":[{"comment":"The phrase 'iven DRK's planar nature' is missing the leading 'G'; please correct it to 'Given DRK's planar nature'.","section":"Abstract"},{"comment":"The column header 'A verage' should be 'Average'.","section":"Tables 1 and S1"},{"comment":"Beyond the numerical issue noted above, the notation 'e^{-32}' is ambiguous; please define what the superscript represents.","section":"Sec. 4.4, Eq. (11)"},{"comment":"The name 'Tank&Temple' is written inconsistently; please use 'Tanks and Temples' throughout.","section":"Sec. S3"},{"comment":"The reference to StopThePop [42] appears with broken citation markers ('6, 4, 5'); please fix the citation formatting.","section":"Sec. 4.4"},{"comment":"The labels 'GS (50)', 'GS (500)', 'GS (5K)', 'DRK (50)' are not defined; state what the numbers denote (e.g., number of primitives).","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is well founded: the paper's own Table S6 shows the method is not robust to pose noise, and Table 3 shows it does not beat 3D-GS (L) on Mip-NeRF360 PSNR. The main text should either address these issues or moderate the claims. The primitive itself is technically sound and interesting; with revised claims and a more balanced discussion, it could be a strong contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is sound and worth building on: DRK replaces Gaussian kernels with a planar radial basis whose angles, lengths, and L1/L2 blend are learnable, plus a piecewise sharpening function. The math checks out—Eq. 6 does reduce to a 2D Gaussian for K=4 orthogonal bases, and the kernel stays simple enough to optimize. The rasterization contributions (ray-plane intersection sorting, polygon-based culling, cache sorting) are reasonable engineering, and the Mesh2DRK conversion is a nice extra. Credit where due: the ablation shows that both the L1/L2 blend and the sharpening matter, and on the authors' own DiverseScenes the primitive count drops dramatically while PSNR and LPIPS improve.\n\nThe soft spots are real and the stress-test note is right. The abstract's \"state-of-the-art performance while dramatically reducing primitive count\" is not true as stated. On Mip-NeRF360 the full DRK uses 952K primitives versus 811K for 3D-GS, with only a small PSNR edge; the sparse variants that cut primitives also cut PSNR. The deeper problem is robustness to camera pose error. Table S6 shows DRK's average PSNR on DiverseScenes falls from 37.58 to 31.59 under 1e-3 pose noise, below 3D-GS at 33.44. That is a 6 dB drop from a tiny perturbation, and Table S5 shows the same effect on Tanks & Temples. This means the advantage is conditional on near-perfect poses, which most real captures do not have. The authors are honest enough to include these tables and attribute the T&T failure to camera error, but the headline should say \"with accurate camera poses,\" not just \"state of the art.\"\n\nMinor but worth noting: the derivation for Eq. 11 is promised in the supplementary but not actually shown there; there are no error bars; and no code or data is released. None of this is fatal, but a serious referee will want the derivation and the claims narrowed.\n\nWho is this for? Anyone working on Gaussian-style scene representations. The kernel is a credible new primitive that likely becomes a baseline in future splatting papers. It deserves peer review, but the authors should be pushed to qualify the SOTA claim and to discuss the pose-sensitivity regime, not just relegate it to a supplementary table.","headline":"A genuinely more expressive splatting kernel that delivers real gains on clean synthetic data, but the unqualified SOTA claim does not survive the paper's own pose-noise experiments.","tokens_in":19846,"tokens_out":1925,"would_cite":true,"duration_ms":19912,"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 a planar deformable radial kernel generalizes Gaussian splatting, letting a single primitive model sharp, non-elliptical shapes and cutting the primitive count by an order of magnitude while improving rendering…","keywords":["3D Gaussian splatting","deformable radial kernel","novel view synthesis","radiance fields","differentiable rendering","shape primitives","kernel culling","pose robustness"],"falsifier":"A decisive test: on a sharp-edged scene with accurately estimated poses and equal primitive counts, DRK must beat Gaussian splatting on PSNR and LPIPS; under $10^{-3}$ pose noise the paper's own table predicts DRK falls below Gaussian splatting, and observing otherwise would overturn its robustness limitation.","tokens_in":18792,"feed_emoji":"🧊","tokens_out":7332,"duration_ms":65784,"temperature":0.7,"pith_summary":"The paper tries to establish that the Gaussian kernel is not the best primitive for radiance-field splatting, and that a planar 'deformable radial kernel' — a shape built from learnable polar radii, angles, a hybrid L1/L2 distance, and a piecewise edge-sharpening map — can represent sharp-edged, non-elliptical scene content with far fewer primitives and better rendering quality. If right, it matters because 3D Gaussian splatting is a dominant real-time rendering method, and DRK offers a drop-in replacement that inherits its pipeline while removing its smoothness bottleneck. The paper reports state-of-the-art PSNR, LPIPS, and SSIM on its new benchmark covering textures, geometry, specular effects, and large scenes, and on a standard real-world unbounded benchmark, with primitive counts reduced by up to an order of magnitude at the sparsest setting. It also shows that the advantage depends on accurate camera poses.","feed_headline":"Deformable kernels render sharp 3D scenes with far fewer primitives","feed_subtitle":"A learnable planar shape primitive beats Gaussian splatting on detail and edge sharpness while using a fraction of the primitives.","key_machinery":"The load-bearing object is the deformable radial kernel: a 2D planar primitive parameterized by center $\\mu$, rotation $q$, K radial lengths $s_k$ and polar angles $\\theta_k$, a curvature blend weight $\\eta \\in (0,1)$, a sharpness parameter $\\tau \\in (-1,1)$, opacity $o$, and spherical harmonics. The kernel function interpolates between adjacent radial bases using cosine weights in the angular coordinate, which is exactly the mechanism that makes a 2D Gaussian a special case; blending in the L1 norm via $\\eta$ gives straight edges, and the piecewise linear map $\\Psi$ decouples edge sharpness from spatial scale. Around this kernel, the paper builds a rasterizer with polygon-based tile culling from the radial endpoints, an 8-entry cache-sorted depth ordering based on ray-plane intersection distance, and a view-dependent low-pass filter adapted from surface splatting.","core_discovery":"The central claim is that replacing the radially symmetric Gaussian's fixed conic falloff with a kernel whose contour is defined by K learnable radial bases — $\\alpha = o\\exp(-\\frac{1}{2}(\\eta r_1^2 + (1-\\eta) r_2^2 / \\bar{s}^2))$ followed by a piecewise sharpening function $\\Psi(g)$ — yields a strictly more expressive planar primitive that still reduces to a 2D Gaussian as a special case when K = 4 with orthogonal axes and paired equal scales. The paper argues this removes three built-in limitations of Gaussians: rotational symmetry, conic L2 boundaries that cannot make straight edges, and the coupling of spatial extent with edge sharpness. The consequence is a splatting method that fits rectangles, triangles, and curved boundaries with single kernels instead of clusters of Gaussians, and the experiments back this with higher PSNR and lower LPIPS on both synthetic and real unbounded scenes.","pith_inferences":["A testable consequence the authors leave implicit: the same sharpness parameters that give DRK its edge fidelity should make it more prone to overfit pose errors, since small misalignments create high-frequency residuals; a joint pose-and-kernel optimization would likely close most of the robustness gap.","The kernel formulation suggests a natural bridge to surface reconstruction: fit DRK's polygon boundary as a mesh proxy, giving a train-free way to extract geometry from splatting.","One could extend DRK to dynamic scenes by making the radial bases and angles time-dependent; the planar form is compatible with existing deformation machinery.","DRK's expressiveness per primitive might also benefit inverse rendering, where a single kernel with a sharp edge can model a material boundary without a separate segmentation."],"forward_implications":["On synthetic and real scenes with accurate poses, DRK matches or exceeds Gaussian-splatting baselines on PSNR, LPIPS, and SSIM while using drastically fewer primitives; at its sparsest setting it uses roughly 42K primitives versus about 336K for standard Gaussian splatting.","Sharp boundaries and straight edges — text, foliage silhouettes, specular rims — become representable by single kernels, so the number of kernels needed to approximate a shape no longer depends on its edge sharpness.","Because DRK is planar and its boundary is a polygon, triangle and polygon meshes can be converted into DRK kernels without training, letting traditional 3D assets be composited into reconstructed scenes.","The paper's own robustness study shows the quality advantage is conditional: under pose noise with standard deviation $10^{-3}$, DRK's PSNR drops from 37.58 to 31.59 while Gaussian splatting only drops from 34.41 to 33.44.","The cache-sorting mechanism, keyed to ray-plane intersection distance, is designed to prevent popping artifacts when overlapping kernels have identical center depths, which is especially relevant for the mesh-conversion application."],"supporting_citations":[{"why":"Baseline Gaussian splatting method that DRK extends; supplies the alpha-blending, densification, and rasterization backbone used for comparison.","marker":"[25]"},{"why":"Planar Gaussian splatting formulation that DRK generalizes from a 2D Gaussian to arbitrary radial contours.","marker":"[20]"},{"why":"Comparison method that tunes Gaussian exponents for sharpness but retains rotational symmetry, motivating DRK's anisotropic bases.","marker":"[18]"},{"why":"Comparison method using hemisphere cutting to handle discontinuities; illustrates why cutting approaches stay tied to Gaussian smoothness.","marker":"[29]"},{"why":"Provides the real-world unbounded scene benchmark on which DRK is compared; supports the claim of strength in well-supervised central regions.","marker":"[2]"},{"why":"Prior work on consistent depth sorting; DRK adapts its insight into an 8-entry cache sorted by ray-plane intersection distance.","marker":"[42]"},{"why":"Gives the surface-splatting low-pass filtering technique that DRK adapts into a view-dependent frequency clamp.","marker":"[4]"},{"why":"Supplies camera poses for the real-scene experiments; the paper attributes DRK's weaker showing on harder scenes to pose error in this estimator.","marker":"[45]"}],"fun_headline_variants":["Deformable radial kernels beat Gaussian splatting","Fewer primitives, sharper edges with deformable kernels","Radial kernel splatting cuts primitive count","Sharp edges from learnable radial bases","Splatting upgrade: flexible kernels outperform Gaussians"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the scene's camera poses are very accurate; the paper's own noise experiments show that with pose error at the level of $10^{-3}$ in the camera parameters, DRK loses its advantage and falls below Gaussian splatting on PSNR.","fun_headline_variants_meta":{"raw":{"variants":["Deformable radial kernels beat Gaussian splatting","Fewer primitives, sharper edges with deformable kernels","Radial kernel splatting cuts primitive count","Sharp edges from learnable radial bases","Splatting upgrade: flexible kernels outperform Gaussians"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1563,"prompt_tokens":896,"completion_tokens":667,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":594}},"tokens_in":512,"tokens_out":667,"duration_ms":5771,"temperature":1.0,"reasoning_tokens":594,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:37:01.767773+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: on a sharp-edged scene with accurately estimated poses and equal primitive counts, DRK must beat Gaussian splatting on PSNR and LPIPS; under $10^{-3}$ pose noise the paper's own table predicts DRK falls below Gaussian splatting, and observing otherwise would overturn its robustness limitation.","supporting_citations":[{"cited_title":"2d gaussian splatting for geometrically ac- curate radiance fields","cited_arxiv_id":null,"evidence_quote":"Planar Gaussian splatting formulation that DRK generalizes from a 2D Gaussian to arbitrary radial contours."},{"cited_title":"Ges: Generalized exponential splatting for efficient radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Comparison method that tunes Gaussian exponents for sharpness but retains rotational symmetry, motivating DRK's anisotropic bases."},{"cited_title":"Mip-NeRF 360: Unbounded anti-aliased neural radiance fields","cited_arxiv_id":null,"evidence_quote":"Provides the real-world unbounded scene benchmark on which DRK is compared; supports the claim of strength in well-supervised central regions."},{"cited_title":"Stopthepop: Sorted gaussian splatting for view-consistent real-time rendering","cited_arxiv_id":null,"evidence_quote":"Prior work on consistent depth sorting; DRK adapts its insight into an 8-entry cache sorted by ray-plane intersection distance."},{"cited_title":"High-quality surface splatting on today’s gpus","cited_arxiv_id":null,"evidence_quote":"Gives the surface-splatting low-pass filtering technique that DRK adapts into a view-dependent frequency clamp."},{"cited_title":"Structure-from-motion revisited","cited_arxiv_id":null,"evidence_quote":"Supplies camera poses for the real-scene experiments; the paper attributes DRK's weaker showing on harder scenes to pose error in this estimator."}],"review_version":1}