{"id":"36d069ce-e0a3-4440-9e73-93dfe2d9b72d","arxiv_id":"2501.11020","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Car-GS combines per-view Gaussians, a separate geometry opacity, and masked normal priors, achieving a mean Chamfer distance of 0.060 on a five-scene subset of 3DRealCar, down from 0.098 for the best prior method.","lead":"Car-GS is a 3D reconstruction method built on Gaussian splatting that targets the reflective paint and transparent glass surfaces of cars, using per-view Gaussians, a separate geometry opacity, and masked normal priors. On five selected real-car scenes, it reports lower Chamfer distance than 2DGS, Gaussian Shader, PGSR, and 3DGS-DR.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's SOTA claim is built on five hand-picked 3DRealCar scenes whose 'reliable GT' selection criterion can directly bias the glass/reflection comparison; without a defined sampling rule or error bars, the 0.060 vs 0.098 CD gap is not yet evidence of general superiority.","rationale":"The central claim is comparative and empirical: Car-GS is state-of-the-art for car reconstruction because its mean CD on five 3DRealCar scenes is 0.060, below Gaussian Shader (0.098), PGSR (0.109), 2DGS (0.135), and 3DGS-DR (0.142). For that comparison to be meaningful, the five scenes must be representative, or at least chosen by a rule independent of the method's expected strengths. The paper's own Section 4.1 reveals the selection rule is GT reliability, and the same paragraph explains why GT reliability is not independent of the phenomenon of interest: 3D scanners struggle with glass. Since the method is specifically designed for glass and reflection, selecting scenes where the scanner GT is reliable could remove the hardest cases or keep unrepresentatively easy ones; either way, the 0.060 number is not a stable estimate of dataset-level performance. This concern is more load-bearing than the rendering-quality overclaim or the missing tau threshold, because those affect secondary claims or reproducibility, whereas scene selection bears directly on whether the headline 'state-of-the-art' claim is true at all. The proposed random-sample or full-dataset rerun with the same protocol would settle it: if the gap persists, the concern is resolved; if it shrinks or reverses, the central claim must be weakened. This matches the reader's weakest assumption, and it supports keeping the CONDITIONAL verdict.","tokens_in":12628,"tokens_out":9360,"duration_ms":99658,"concrete_test":"Run Car-GS and Gaussian Shader (the strongest baseline in Table 1) on a pre-registered random sample of at least 20 additional 3DRealCar scenes that have any scanner GT, using the same TSDF protocol, and report per-scene CD, mean CD, and the number of scenes where Car-GS wins. If the mean CD advantage over Gaussian Shader is substantially smaller than the Table 1 gap (0.060 vs 0.098) or Car-GS wins on fewer than 15 of 20 scenes, the five-scene result is not a reliable basis for the SOTA claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 says: 'Due to the inherent challenges in accurately capturing glass surfaces with a 3D scanner, we carefully selected five scenes with reliable ground truth data to ensure robustness and fairness.' This sentence is the load-bearing step for the paper's central empirical claim. Table 1 reports a mean Chamfer Distance of 0.060 on exactly those five scenes, and all three contributions (VDG, LHO, QSM) target reflective and transparent surfaces. Selecting scenes by GT reliability can select on the outcome: if scanner confidence is lower where glass reflections are hardest, the chosen subset may omit the cases where baselines fail most, making Car-GS look better than it is on the full distribution. The paper gives no sampling rule, no number of scenes considered before selection, no GT-confidence score, no error bars over scenes or random seeds, and no full-dataset or random-subset comparison. Table 3's full-model accuracy/F1 of 0.636/0.664 exactly match Scene 4 in Table 1, suggesting the ablation may be on a single scene, so even the component analysis does not establish cross-scene robustness. Without an a priori inclusion criterion, the SOTA claim is not separable from selection bias.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Car-GS, a 2DGS-based method for 3D car surface reconstruction that addresses reflective and transparent surfaces through three components: view-dependent Gaussians (VDG) that absorb specular highlights per view, a learnable hybrid opacity (LHO) that separates appearance and geometry opacities, and a quality-aware supervision module (QSM) that weighs normal priors from a pretrained model by the angle between the ray and the Gaussian normal. The central empirical claim is state-of-the-art car surface reconstruction on five 3DRealCar scenes, with mean Chamfer Distance 0.060 versus 0.098 for Gaussian Shader and 0.109 for PGSR (Table 1), plus qualitatively better geometry on Ref-NeRF Real scenes.","tokens_in":12965,"tokens_out":2244,"duration_ms":25504,"significance":"If the reported geometry gains hold across the full 3DRealCar distribution, the paper would be a useful contribution to reflective/transparent surface reconstruction, a known weakness of Gaussian splatting methods. The design is reasonably grounded: the VDG/LHO/QSM components are motivated by physically plausible observations about specular highlights and ray termination on glass, and the geometry evaluation uses externally acquired high-precision scanner meshes rather than self-derived targets. The paper also provides a practical training-time advantage (20 minutes versus 1h40m for Gaussian Shader). However, the evidence for the headline SOTA claim currently rests on five hand-picked scenes with no selection rule, no error bars, and no full-dataset or random-subset validation, and the paper's own Table 2 contradicts the accompanying rendering-superiority claim.","major_comments":[{"comment":"The 'carefully selected five scenes with reliable ground truth data' sentence is load-bearing for the central SOTA claim, but no selection criterion is given. Since all three contributions target exactly the reflective/transparent regions where scanner GT is hardest to obtain, a selection rule based on GT reliability can select on the outcome, removing the cases where baselines fail most. The paper should state the total number of scenes inspected, the numeric GT-confidence score used for selection, and ideally report results on a random subset and/or the full dataset. Without this, the 0.060 versus 0.098 Chamfer Distance gap is not separable from selection bias.","section":"Section 4.1, Table 1"},{"comment":"The text states 'our approach also demonstrates superior rendering performance,' but Table 2 contradicts this: on 3DRealCar, Ours has PSNR 23.982 versus PGSR 25.818, SSIM 0.803 versus 0.876, and LPIPS 0.258 versus 0.122; on Ref-NeRF, Ours is also not best on any metric. The rendering claim is therefore false as stated. The authors should either remove the rendering-superiority claim or reframe it as competitive rendering, and should not use 'superior' in the abstract or Section 4.3 to describe rendering quality.","section":"Section 4.3, Table 2"},{"comment":"The quality-aware supervision module depends on the threshold tau, but tau is never specified in the paper or in Section 4.2's implementation details. Since QSM is one of the three contributions and the loss mask in Fig. 4 is directly determined by tau, the paper should report the value used, any normalization convention for the angle, and a sensitivity study over tau. Without this, the adaptive supervision is not reproducible.","section":"Section 3.3, Eq. (10)"},{"comment":"The ablation table reports a single set of numbers with no indication of which scene(s) were used. The full-model CD 0.060, Accuracy 0.636, and F1 0.664 exactly match Scene 4 in Table 1, strongly suggesting the ablation was run on one scene only. If that is the case, the paper should say so explicitly and provide per-scene ablation results or at least means and variances over multiple scenes; otherwise the component analysis cannot support a cross-scene robustness claim.","section":"Section 4.4, Table 3"}],"minor_comments":[{"comment":"The scale s and offset o for aligning monocular depth with COLMAP are said to be 'obtained following the method in 3DGS [17]', but 3DGS does not perform monocular depth alignment. Please cite the actual alignment procedure or describe the fitting objective.","section":"Section 3.1, Eq. (3)"},{"comment":"The module is called Quality-aware Supervised Module (QSM) in Section 3.3 and the abstract, but the Table 3 caption abbreviates it as QAS. Please unify the abbreviation throughout.","section":"Section 3.3 and Table 3 caption"},{"comment":"The sentence 'leverages ground truth normals from a pre-trained normal estimation model' is misleading because these are pseudo labels, not ground truth. Please replace 'ground truth' with 'pseudo' or 'predicted' normals.","section":"Section 3.3, text near Eq. (9)"},{"comment":"The Ref-NeRF Real dataset has no ground truth geometry, so the comparison there is only visual. Please state this limitation explicitly and avoid wording like 'consistently yields geometrically accurate surface reconstructions' without quantitative support on that dataset.","section":"Section 4.1"},{"comment":"The caption says 'black indicates a value of 0, while white indicates a value of 1' for the mask, but Eq. (10) sets wn=0 for the orthogonal regions that are described as errors. Please confirm whether these regions are black in Fig. 4(c) and clarify the color mapping.","section":"Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The core geometry contribution is plausible and the external scanner GT is a genuine strength, but the evaluation protocol is currently too fragile to support the SOTA claim: the five-scene selection rule is undefined, and the paper's own rendering table contradicts a prominent claim. I would encourage the editor to request a revised version that addresses the selection-bias concern with a defined inclusion protocol or full-dataset results, specifies tau and other hyperparameters, and corrects the rendering claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Car-GS is a reasonable engineering paper. The three components are each motivated and the integrated system reports a clean win on five 3DRealCar scenes (CD 0.060 vs 0.098 for Gaussian Shader). That result is internally consistent with the ablation, and the use of scanner-derived ground truth is a real strength. The paper is honest about drawing on prior work; the novelty is the combination, not any single idea.\n\nThe soft spots are mostly in the evaluation and in one overclaim. The five scenes are described as 'carefully selected with reliable ground truth,' which is exactly the kind of selection that can bias a comparison on glass and reflections. There are no error bars, no per-scene variance, no random-subset or full-dataset check. The threshold tau in Eq. (10) is never specified or ablated. Table 3's numbers match Scene 4 exactly, which suggests the ablation ran on a single scene, though the text implies otherwise. And the abstract/conclusion claim of superior rendering is contradicted by Table 2, where PGSR is clearly better on PSNR/SSIM/LPIPS on 3DRealCar. The limitation section is also empty of limitations.\n\nNone of this breaks the core idea. The geometry claim on the selected scenes is plausible, and the method is worth testing. But as submitted, the SOTA generalization is not supported. I'd want a defined scene-selection rule, results on a random subset or the full dataset, error bars, a specified tau, and a correction of the rendering claim before I'd trust the headline.\n\nThis is a paper for people working on GS-based surface reconstruction for cars and reflective/transparent objects. It deserves a serious referee, not a desk reject, but it needs a revision that closes the evaluation gaps.","headline":"Plausible integrated method for reflective/transparent car surfaces, but the SOTA claim rests on five hand-picked scenes and one overclaim that contradicts the paper's own table.","tokens_in":13451,"tokens_out":2057,"would_cite":true,"duration_ms":22772,"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":"Car-GS reconstructs reflective car surfaces with a mean Chamfer distance of 0.060 on five 3DRealCar scenes by splitting appearance from geometry and masking unreliable normal priors.","keywords":["3D Gaussian splatting","surface reconstruction","reflective surfaces","transparent surfaces","view-dependent Gaussians","learnable hybrid opacity","normal prior supervision","3DRealCar"],"falsifier":"Run the same training and evaluation pipeline on the full 3DRealCar dataset, or on a randomly drawn held-out set of scenes, and report mean Chamfer Distance with per-scene confidence intervals; if the advantage over Gaussian Shader at 0.098 and PGSR at 0.109 shrinks to noise outside the five selected scenes, the state-of-the-art generalization claim is falsified.","tokens_in":1763,"feed_emoji":"🚗","tokens_out":1801,"duration_ms":90872,"temperature":0.7,"pith_summary":"The paper tries to establish that 3D car surfaces, which are notoriously hard to reconstruct because of glossy paint and transparent glass, can be recovered accurately by extending 2D Gaussian Splatting with three targeted mechanisms. The stakes matter because current Gaussian-splatting baselines confuse view-dependent reflections and transparent surfaces, producing noisy or missing geometry on cars, and accurate car meshes are directly useful for autonomous-driving simulation, AR/VR, and games. The paper reports the best surface-reconstruction numbers on its five-scene subset of 3DRealCar, with mean Chamfer Distance 0.060, and attributes the gain to view-dependent Gaussians for reflections, a learnable hybrid opacity that decouples geometry from appearance, and a quality-aware normal-prior mask.","feed_headline":"Gaussian splatting now handles car glass and gloss","feed_subtitle":"Car-GS separates reflections and transparency, reaching mean Chamfer distance 0.060 on five 3DRealCar scenes.","key_machinery":"The load-bearing objects are the view-dependent Gaussian (VDG) set, the learnable hybrid opacity (LHO), and the angle-gated normal supervision. VDGs are extra Gaussian primitives initialized from monocular depth and aligned to structure-from-motion, one set per training view, regularized toward low opacity, and omitted at mesh extraction; they isolate specular highlights so the view-shared Gaussians learn only view-independent appearance. LHO gives each shared Gaussian two opacity values, $\\alpha$ for RGB and $\\alpha_{\\mathrm{geo}}$ for geometry, rendered through the same 2DGS volume-splatting equation but with a consistency loss $L_{\\mathrm{lho}}$ that keeps the two close enough to avoid degenerate transparent points. The quality-aware module computes the supervision weight as $w_n = 1$ if $\\theta > \\tau$ and $0$ otherwise, where $\\theta$ is the angle between the Gaussian normal and the pixel ray, using pretrained normal predictions as pseudo labels only where they are reliable.","core_discovery":"The central claim is that a Gaussian-splatting reconstruction pipeline can separate the three failure modes of car surfaces and handle each explicitly. View-dependent Gaussians are trained per viewpoint to absorb specular highlights and are discarded at test time. A learnable hybrid opacity gives each 2D Gaussian a geometry-specific opacity for depth and normal rendering distinct from its RGB opacity, so transparent glass can be geometrically solid while visually transparent. A quality-aware supervision module gates normal-prior supervision by the angle between each Gaussian normal and the viewing ray, zeroing the weight when the view is nearly orthogonal to glass. The paper's evidence is the five-scene 3DRealCar evaluation, where Car-GS reaches mean Chamfer Distance 0.060, Accuracy 0.751, and F1 0.764, beating Gaussian Shader at 0.098 CD and PGSR at 0.109 CD, plus ablations showing each module contributes to the gain.","pith_inferences":["Extension: the appearance/geometry opacity split is not car-specific; the same learnable hybrid opacity could be tested on other transparent objects such as bottles or building windows, where depth and RGB rays also diverge.","Extension: because view-dependent Gaussians are discarded at evaluation, Car-GS removes specular reflections rather than explaining them, so a follow-up that retains the VDG would be needed for relighting or material editing, which the paper does not address.","Extension: the angle-based normal mask could be applied to any normal-prior-guided Gaussian reconstruction, not just cars; a cheap transfer test would run the same quality-aware supervision on a general object dataset and compare surface error against methods with unmasked priors.","Extension: a testable prediction is that the largest gains over baselines appear exactly in near-orthogonal glass views; a per-view error analysis over the five scenes would confirm or refute the mechanism claimed for the quality-aware supervision module."],"forward_implications":["On the five 3DRealCar scenes, Car-GS reaches a mean Chamfer Distance of 0.060, beating Gaussian Shader at 0.098 and PGSR at 0.109, which is the concrete basis for its state-of-the-art surface-reconstruction claim.","Because view-dependent Gaussians are discarded at mesh extraction, the reflection-handling machinery adds no cost at inference; the reported training time is 20 minutes per scene, near the 15 minutes of plain 2DGS and much less than Gaussian Shader's 1h40m.","The learnable hybrid opacity implies that depth and normal maps can terminate at a transparent surface while RGB still sees through it, directly targeting glass-penetration artifacts in windshields and windows.","With quality-aware normal supervision, a general-purpose pretrained normal prior can be used without letting its errors corrupt geometry, particularly in near-orthogonal glass views where the paper says reconstruction errors concentrate.","The method also produces usable novel-view synthesis on 3DRealCar and Ref-NeRF Real, so the geometry-focused changes do not eliminate appearance reconstruction."],"supporting_citations":[{"why":"Supplies the base 3D Gaussian Splatting rendering, initialization scale and offset alignment, and RGB loss that Car-GS builds on.","marker":"[17]"},{"why":"Supplies the 2D Gaussian primitive representation, the depth-normal consistency loss, and the TSDF surface-extraction recipe Car-GS inherits.","marker":"[14]"},{"why":"Supplies the 3DRealCar dataset and scanner-derived ground-truth meshes used for all surface-reconstruction evaluation.","marker":"[9]"},{"why":"Supplies the pretrained normal estimates used as pseudo labels in the quality-aware supervision module.","marker":"[43]"},{"why":"Supplies the Ref-NeRF Real dataset and the reflective-surface reconstruction problem against which Car-GS is compared.","marker":"[31]"},{"why":"Supplies the strongest reflective-surface baseline, Gaussian Shader, that Car-GS must beat for its state-of-the-art claim.","marker":"[16]"},{"why":"Supplies the monocular depth estimates that initialize the view-dependent Gaussians in textureless reflection regions.","marker":"[39]"},{"why":"Supplies the instance segmentation mask used to sample pixels for view-dependent Gaussian initialization.","marker":"[27]"}],"fun_headline_variants":["Car-GS: Splatting that separates glass from geometry","Gaussian splatting now handles car reflections and transparency","New method fixes car reconstruction by decoupling RGB and depth","Car-GS: Adaptive normal priors for transparent car surfaces"],"cache_read_input_tokens":15488,"weakest_assumption_plain":"The load-bearing premise is that the five 3DRealCar scenes chosen for reliable scanner ground truth represent how hard car surfaces are overall; if those scenes are unrepresentatively easy or favorable to this method, the reported state-of-the-art numbers do not generalize to the rest of the 2,500-vehicle dataset.","fun_headline_variants_meta":{"raw":{"variants":["Car-GS: Splatting that separates glass from geometry","Gaussian splatting now handles car reflections and transparency","New method fixes car reconstruction by decoupling RGB and depth","Car-GS: Adaptive normal priors for transparent car surfaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000276,"raw_usage":{"total_tokens":1659,"prompt_tokens":971,"completion_tokens":688,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":619}},"tokens_in":587,"tokens_out":688,"duration_ms":7563,"temperature":1.0,"reasoning_tokens":619,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:42:38.754259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same training and evaluation pipeline on the full 3DRealCar dataset, or on a randomly drawn held-out set of scenes, and report mean Chamfer Distance with per-scene confidence intervals; if the advantage over Gaussian Shader at 0.098 and PGSR at 0.109 shrinks to noise outside the five selected scenes, the state-of-the-art generalization claim is falsified.","supporting_citations":[{"cited_title":"2d gaussian splatting for geometrically accu- rate radiance fields","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D Gaussian primitive representation, the depth-normal consistency loss, and the TSDF surface-extraction recipe Car-GS inherits."},{"cited_title":"Stablenormal: Reducing diffusion variance for stable and sharp normal","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained normal estimates used as pseudo labels in the quality-aware supervision module."},{"cited_title":"Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields","cited_arxiv_id":null,"evidence_quote":"Supplies the Ref-NeRF Real dataset and the reflective-surface reconstruction problem against which Car-GS is compared."},{"cited_title":"Gaussian- shader: 3d gaussian splatting with shading functions for re- flective surfaces","cited_arxiv_id":null,"evidence_quote":"Supplies the strongest reflective-surface baseline, Gaussian Shader, that Car-GS must beat for its state-of-the-art claim."}],"review_version":1}