{"id":"61ff21cf-b69f-44d1-b98d-73881140b889","arxiv_id":"2412.10084","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"By encoding angular appearance in low-resolution spherical-harmonic light field probes instead of full-resolution per-voxel color features, ProbeSDF makes SDF-based surface reconstruction faster and more accurate on four benchmarks.","lead":"ProbeSDF separates the two things a 3D reconstruction model must learn: fine surface detail lives in a high-resolution grid, while lighting and view-dependent shading live in a coarser grid of light field probes. On four benchmark datasets the method rebuilds objects and people in one to four minutes with equal or better geometry and image quality than prior pipelines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation asymmetries, not the probe approximation, are the load-bearing risk: per-camera bias vectors and resolution mismatches confound the claimed superiority on DTU/BMVS and ActorsHQ.","rationale":"Good-faith reading: ProbeSDF is a systems paper with a clear hypothesis—angular radiance is spatially smooth and can be stored on coarse probes—and the implementation is careful: fused kernel, ablations, per-scene tables, and a Limitations section that names the shadow-artifact failure mode on ActorsHQ. The geometric results on MVMannequins and DTU are substantial and are measured against external ground truth, so I do not see an internal inconsistency or a circular argument. However, the headline claim is comparative ('superseding'), and the comparisons contain three uncontrolled asymmetries that the paper itself discloses: per-camera bias vectors on DTU/BMVS, resolution mismatch on ActorsHQ, and train-set PSNR. These are not mere presentational issues; they can change the ranking. The per-camera bias is particularly load-bearing because it adds parameters that only ProbeSDF uses, and the photometric loss is the main driver of SDF updates. The reader's chosen weakest_assumption (probe resolution) is a real limitation and is honestly acknowledged, but it is less central to the benchmark claim: on the four datasets the assumption apparently holds well enough to produce the reported geometry, and the paper's own failure analysis is consistent. The evaluation confounds are the place where the central claim could actually be false even if the method works as described. Hence the verdict remains CONDITIONAL: accept only after a controlled re-run. I partially agree with the reader: they identified these protocol asymmetries in their rationale, but their weakest_assumption field points to the probe approximation rather than the evaluation control. My read is that the evaluation control is the more load-bearing issue.","tokens_in":20470,"tokens_out":6117,"duration_ms":61440,"concrete_test":"Retrain the baselines on DTU and BlendedMVS with per-camera bias vectors of the same form as ProbeSDF's (e.g., a per-camera learnable offset added to the color before the final sigmoid), and retrain ProbeSDF without them, keeping all other protocol choices fixed; then compare mean chamfer and PSNR. Separately, on ActorsHQ, train ProbeSDF and Voxurf at r/2 with at least three seeds and report mean±std PSNR; if ProbeSDF's advantage over Voxurf at matched resolution is within run-to-run noise, the ActorsHQ part of the 'superseding' claim is unsupported. If the DTU/BMVS chamfer advantage persists without bias vectors, the geometry claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that ProbeSDF supersedes prior methods on four benchmarks—rests on the comparability of the reported evaluations, and that comparability is not established. (1) On DTU and BlendedMVS, ProbeSDF optionally trains per-camera bias vectors in the MLP (Sec. 5, Tables 3 and 4, ✓ marks), while Voxurf, NeuS2, and 2DGS are not given this mechanism. These biases can absorb per-camera exposure/color inconsistencies, inflating PSNR and potentially relaxing the photometric constraint that drives SDF optimization; the PSNR gaps (37.03/37.74 vs 37.08/36.00/36.03 on DTU; 35.19/35.89 vs 35.11/33.62 on BMVS) are therefore not apples-to-apples. (2) On ActorsHQ, Voxurf and NeuS2 are trained at half resolution because they 'ran into issues' at full resolution, while ProbeSDF is trained at full resolution; the headline PSNR advantage (37.48 vs 36.56) is mostly a resolution advantage, and the matched-resolution ProbeSDF r/2 result (36.62) exceeds Voxurf r/2 by only 0.06 dB. (3) PSNR is computed on the training images ('all available images were given as input and used for evaluation'), so the image-metric claim measures training fit, not generalization. The chamfer results on MVMannequins and DTU are evaluated against external ground truth and are less affected, but the 'unprecedented performance combination' across all four datasets is not yet supported. A controlled re-run with matched capabilities is required before the superiority claim can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ProbeSDF proposes a reparameterization of the appearance model for SDF-based multi-view reconstruction: spatial features are stored in a high-resolution sparse voxel grid, while angular features are stored in a coarser grid of light-field probes using low-order spherical harmonics, decoded by a tiny MLP that also receives powers of (1 - n·v) as a learned Fresnel proxy. The method is evaluated on MVMannequins, ActorsHQ, DTU, and BlendedMVS against Voxurf, NeuS2, MMH, and 2DGS, with per-scene tables, ablations over (ns, na, l), probe smoothing, the Fresnel term, and qualitative comparisons. The central claim is that this decoupling yields superior geometry and image metrics on all four benchmarks while providing large training and rendering speedups.","tokens_in":20730,"tokens_out":6351,"duration_ms":71737,"significance":"If the reported comparisons are accepted at face value, the contribution is significant and practical: the probe-based angular parametrization is conceptually clean, the MLP is dramatically smaller, training is minutes rather than tens of minutes, and rendering reaches hundreds of Hz. The paper has several concrete strengths: the baselines are re-trained under a stated common protocol, per-scene numbers are provided for all datasets, the ablations cover the main knobs of the parametrization, and source code is promised. The geometric results against external ground truth (chamfer on MVMannequins, DTU, BlendedMVS) appear substantially better than several baselines, which is the most trustworthy part of the evidence. However, the headline claim of superiority rests on evaluation asymmetries that are not yet resolved, so the contribution is promising but not yet established at the level claimed.","major_comments":[{"comment":"The DTU and BlendedMVS comparisons are confounded by per-camera bias vectors. The text states that 'some datasets may contain inconsistent shadows or large exposure changes (DTU and BMVS) that are hard to model by our representation so we optionally train per-camera bias vectors in the MLP,' and Tables 3 and 4 mark the ProbeSDF rows with a checkmark for this option. Voxurf, NeuS2, and 2DGS are not given an equivalent per-camera or per-image adjustment. These bias vectors can absorb per-camera exposure/color differences, which directly inflates the image metric and also changes the photometric gradients that drive SDF optimization, so both the PSNR gaps (37.03/37.74 versus 37.08/36.00/36.03 on DTU; 35.19/35.89 versus 35.11/33.62 on BlendedMVS) and the chamfer comparisons are not apples-to-apples. A controlled comparison that either trains the same bias mechanism for the baselines or removes it from ProbeSDF is required before the claimed superiority on these two benchmarks can be accepted.","section":"Sec. 5, Tables 3 and 4; Sec. 5 text"},{"comment":"The ActorsHQ comparison is asymmetric in input resolution. Voxurf and NeuS2 are trained at half resolution 'since we ran into issues with full resolution images,' while ProbeSDF is trained at full resolution. The headline PSNR of 37.48 dB is at r/1; at the matched resolution r/2, ProbeSDF obtains 36.62 dB versus 36.56 dB for Voxurf, a difference of only 0.06 dB. The statement that ProbeSDF 'also outperform[s] both when training at the same resolution' is technically true but not representative of the claimed margin, and the full-resolution comparison for the baselines is missing. This is a load-bearing issue because the resolution advantage is the main source of the reported 0.92 dB gap. A fair conclusion requires either full-resolution baseline runs or a presentation in which the matched-resolution numbers are the primary head-to-head comparison.","section":"Sec. 5, Table 2 and ActorsHQ paragraph"},{"comment":"The image-metric claim is computed on the training images, not on held-out or novel views. The protocol says 'all available images were given as input and used for evaluation,' and PSNR is computed within the silhouette region. Consequently, the reported PSNR measures training-set reconstruction fidelity rather than generalization, and a model with extra per-camera parameters can increase this number without any improvement in geometry or in the rendering of unseen viewpoints. The geometric chamfer results against external scanners are less affected, but the paper should explicitly qualify the image-metric superiority claim as a training-fit measure, and ideally add a held-out-view PSNR evaluation or at least an analysis showing that the ranking is stable under held-out views.","section":"Sec. 5, Protocol paragraph"}],"minor_comments":[{"comment":"The phrase 'convolved by a53 gaussian kernel' appears to be a missing-formatting error; it should read 'a 5×3 Gaussian kernel' or 'a 5-tap Gaussian kernel' as appropriate.","section":"Sec. 3.3"},{"comment":"The resolution '20482' should be typed as '2048×2048' or '2048²' to avoid a formatting artifact.","section":"Sec. 5, MVMannequins"},{"comment":"Reference [28] lists the last author as 'Franco Jean-Sébastien'; this should be normalized to 'Jean-Sébastien Franco' to match the author list of the present paper and standard citation style.","section":"References"},{"comment":"The phrase 'superseding both 3D and image metric performances with four popular benchmarks' is ambiguous and slightly overreaching: on ActorsHQ no independent geometric metric is reported, so the claim of both 3D and image superiority does not apply to all four benchmarks individually.","section":"Abstract and Introduction"},{"comment":"The ablations cover the feature dimensionalities, the SH order, probe smoothing, and the Fresnel input, but not the probe grid resolution (the fixed 1/16th ratio). Given the central modeling assumption, an ablation varying this ratio would strengthen the paper.","section":"Sec. 5, ablations"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically well-executed and the idea is novel enough for a strong venue, but the main claims are currently supported by comparisons that give ProbeSDF additional capabilities or resolution that the baselines do not have. The required fixes are within the scope of a revision: re-run the baselines with matched capabilities (per-camera bias, full resolution), or clearly reposition the claims. If the authors provide matched comparisons and the ranking survives, I would support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on ProbeSDF (arXiv:2412.10084). The core idea is genuinely new and worth building on: instead of co-locating spatial and angular color features at the same resolution, they store angular appearance as spherical-harmonic 'light field probes' on a grid 1/16th of the SDF grid's resolution, sample them along the reflected vector, and decode with a two-layer MLP fed by Fresnel-like powers of (1−n·v). It is a clean, physically motivated refactor of the appearance model, and the efficiency numbers are real: minutes of training, 200–400 Hz rendering, geometry at or above re-trained baselines on four datasets. The evaluation is extensive—per-scene tables, ablations over (ns, na, l), probe smoothing and the Fresnel term—and the Limitations section honestly concedes the failure modes (close lighting, self-shadows, the ActorsHQ latitude-coverage artifacts). Code is released. The sparse-grid backbone comes from their own MMH, but the comparison is carried by external baselines and external ground-truth geometry, so I do not see circularity.\n\nWhere I push back, and here I agree with the stress-test: the evaluation asymmetries are the load-bearing risk, not the probe approximation. On DTU and BMVS, ProbeSDF trains per-camera bias vectors and the baselines do not get them; on ActorsHQ, Voxurf and NeuS2 run at half resolution, and the matched-resolution gap is 0.06 dB (36.62 vs 36.56). PSNR is reported on the training images, so the image-metric claim is a training-fit claim. The abstract's 'unprecedented performance combination' overshoots what the tables support, given those asymmetries. The authors disclose all of this themselves, which is to their credit, but the superiority claim needs a controlled re-run with matched capabilities before it is treated as state of the art.\n\nThe geometry conclusions are less exposed: chamfer on MVMannequins and DTU uses external ground truth, and the no-bias DTU variant (0.71 mm) roughly matches Voxurf (0.73 mm). So the central architectural claim—decoupled low-resolution angular features are a viable and much faster drop-in appearance model—holds up. The soft spots are in the comparative protocol, not the design.\n\nThis deserves a serious referee. Send it to review with the request that the comparability issues be fixed: give baselines the same bias mechanism or drop it, report test-set or held-out metrics, and add multiple runs. The idea will be built on either way.","headline":"Genuinely new decoupled appearance model with real speed and credible geometry, but the 'supersedes all baselines' claim is constrained by per-camera biases and resolution asymmetries in the evaluation.","tokens_in":21379,"tokens_out":3772,"would_cite":true,"duration_ms":37681,"reading_group":"yes","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 co-located spatial-and-directional appearance decoders are the bottleneck in SDF-based surface reconstruction, and that replacing them with separate fine spatial and coarse angular probe grids yields faster training…","keywords":["neural surface reconstruction","signed distance function","light field probes","spherical harmonics","volume rendering","multi-view 3D reconstruction","real-time rendering","radiance field decoupling"],"falsifier":"Train ProbeSDF on a synthetic glossy object illuminated by a point light placed close to the surface (distance comparable to object size) and compare against a second, distant-light condition while keeping the SDF grid resolution fixed; if chamfer error grows substantially and PSNR plateaus in the near-light case relative to a co-located decoder, the low-frequency probe assumption is falsified. A more direct check is to compare the interpolated probe-radiance field $F_a(x,r)$ at a surface point against a dense ground-truth angular radiance sampled from the renderer; the paper's assumption predicts that the residual rises as light distance shrinks.","tokens_in":20133,"feed_emoji":"⚡","tokens_out":6823,"duration_ms":68472,"temperature":0.7,"pith_summary":"SDF-based neural surface reconstruction normally stores color and view-direction information together in one high-capacity decoder. This paper claims that the two components have very different spatial bandwidth: surface texture varies fast, while lighting, and therefore angular radiance, varies slowly because light sources are usually far away. It therefore separates them, encoding spatial appearance in a fine voxel grid and angular appearance in a coarse grid of spherical-harmonic 'light field probes' at one sixteenth of the voxel resolution, with a tiny two-layer MLP and a polynomial Fresnel term as the only decoder. On human and object benchmarks (MVMannequins, ActorsHQ, DTU, BlendedMVS) the paper reports that this minimal refactor beats re-trained baselines on chamfer distance and PSNR while training in 1–4 minutes and rendering at 200–400 Hz. A sympathetic reader would take away that the co-located decoder is not necessary and that separating spatial and angular bandwidth is a cheap, general improvement.","feed_headline":"Low-res light probes sharpen and speed up SDF reconstruction","feed_subtitle":"Separating spatial and angular radiance beats co-located decoders on four benchmarks while training in 1–4 minutes.","key_machinery":"Light-field probes: a coarse grid (1/16 of voxel resolution) whose nodes store spherical-harmonic coefficient vectors $b_{ij}$; at a shading point the probe features are trilinearly interpolated (Eqs. 6–7) and the SH basis is evaluated at the reflected vector $r$, giving $F_a(x,r)$. Spatial features use a planar factorization $F_s = F_x\\otimes F_y\\otimes F_z$ per $16^3$ tile. A two-hidden-layer 32-neuron MLP decodes $F_s$, $F_a$, and the powers $(1-n\\cdot v)^k$ (k=0..5), approximating the Fresnel term, into color. Everything—probe interpolation, feature lookup, and the MLP call—is fused into one single CUDA kernel, and the probes add only 1/6 of the spatial-feature memory (8×16 coefficients vs 3×16×16 per tile at $l=4$).","core_discovery":"The central discovery is Eq. 4: $C = \\mathrm{MLP}(F_s(x), F_a(x,r), (1-n\\cdot v)^0,\\dots,(1-n\\cdot v)^5)$, where the spatial features $F_s(x)$ come from a high-resolution sparse voxel grid and the angular features $F_a(x,r)$ are computed by trilinearly interpolating spherical-harmonic coefficients stored in a coarse probe grid and evaluating them at the reflected direction $r$ (Eqs. 5–7). This decoupling rests on the claim that angular radiance is nearly parallax-invariant because lighting is mostly distant, so a low spatial frequency probe grid suffices. The paper's tables show this parametrization consistently improves chamfer distance and PSNR over MMH, Voxurf, NeuS2, and 2DGS on MVMannequins (1.04 mm), DTU (0.68 mm), BlendedMVS (2.21 mm), and ActorsHQ (37.48 dB), while cutting training time by roughly 3–15x and reaching 200–400 Hz rendering.","pith_inferences":["The paper's own bandwidth argument suggests a test: for scenes with near-field lighting, the probe grid at 1/16 resolution should become the limiting factor, and one could verify this by training on synthetic scenes with a point light at decreasing distances and measuring chamfer error against probe spacing.","Because probes and geometry are optimized jointly but stored separately, the angular grid is a natural handle for relighting or lighting transfer: after training, editing probe coefficients (or fine-tuning them to a new environment) could update appearance without retraining geometry, which the paper lists as future work.","The same decoupling could extend to temporal reconstruction: since lighting is smooth in space, probe coefficients might be shared or tracked across frames, a direction the paper mentions but does not explore.","The ablation showing that $l=1$ (constant SH) degrades chamfer more than PSNR suggests geometry itself is regularized by the angular shading model, implying that probe resolution and SH order are geometric hyperparameters, not just appearance hyperparameters."],"forward_implications":["The probe-based appearance model can serve as a drop-in replacement for the co-located decoders in SDF pipelines, since the SDF opacity and regularization structure (NeuS equations, eikonal, smoothness losses) is unchanged.","Training times drop to 1–4 minutes on a single RTX A6000 for both object and human scenes, putting millimetric reconstruction within interactive capture workflows.","Rendering at 200–400 Hz makes real-time inspection and rendering of reconstructed surfaces practical with a roughly 30–232 MB model.","Raising the SH order $l$ from 1 to 4 or increasing $n_s,n_a$ systematically improves PSNR and chamfer, so specularity is controllable with a single parameter knob, as shown in the ablations.","Per-camera bias vectors absorb inconsistent exposure and shadows on DTU and BlendedMVS, extending the method to casually lit multi-view captures."],"supporting_citations":[{"why":"Provides the sparse voxel grid implementation and fusion kernel ProbeSDF modifies, plus the MVMannequins dataset and the primary speed/accuracy baseline MMH.","marker":"[28]"},{"why":"Supplies the SDF-to-opacity volume rendering equations (Eqs. 8–10) that ProbeSDF retains for differentiable surface rendering.","marker":"[31]"},{"why":"Provides the dense voxel SDF baseline Voxurf and the Gaussian smoothing regularization on raw SDF values that ProbeSDF adapts.","marker":"[34]"},{"why":"Is the hash-grid baseline NeuS2 and the comparison point for training speed and grid-aligned artifacts.","marker":"[32]"},{"why":"Contributes the spherical-harmonic angular encoding and the coarse-to-fine training observation used to schedule the SH order $l$.","marker":"[14]"},{"why":"Is the DTU baseline 2DGS whose no-mask protocol and half-resolution evaluation ProbeSDF matches for comparison.","marker":"[7]"},{"why":"Motivates conditioning appearance on the reflected vector rather than the view vector, which ProbeSDF adopts for probe sampling.","marker":"[29]"},{"why":"Is the conceptual origin of caching lighting at sparse 3D locations as light field probes, which ProbeSDF turns into an optimizable representation.","marker":"[25]"}],"fun_headline_variants":["Coarse light probes speed SDF training 15x and sharpen shapes","ProbeSDF: Split radiance, cut training time, boost accuracy","Low-res angular grids make SDF reconstructions faster and better","Decoupling radiance in SDF yields real-time rendering and sharper meshes","ProbeSDF: Tiny probe grids outperform big decoders on 4 benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that angular radiance varies slowly enough in space to be represented by probes stored at one sixteenth of the voxel resolution with at most fourth-order spherical harmonics, which the paper states directly as the assumption that 'the lights are far enough, so that probes at a low spatial frequency can explain the parallax.' If a scene contains close light sources, sharp self-shadows, or strong local inter-reflections, the probe grid cannot explain the images and the shape-radiance optimization will push geometry off.","fun_headline_variants_meta":{"raw":{"variants":["Coarse light probes speed SDF training 15x and sharpen shapes","ProbeSDF: Split radiance, cut training time, boost accuracy","Low-res angular grids make SDF reconstructions faster and better","Decoupling radiance in SDF yields real-time rendering and sharper meshes","ProbeSDF: Tiny probe grids outperform big decoders on 4 benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1318,"prompt_tokens":930,"completion_tokens":388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":290}},"tokens_in":546,"tokens_out":388,"duration_ms":4947,"temperature":1.0,"reasoning_tokens":290,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:23:19.502541+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ProbeSDF on a synthetic glossy object illuminated by a point light placed close to the surface (distance comparable to object size) and compare against a second, distant-light condition while keeping the SDF grid resolution fixed; if chamfer error grows substantially and PSNR plateaus in the near-light case relative to a co-located decoder, the low-frequency probe assumption is falsified. A more direct check is to compare the interpolated probe-radiance field $F_a(x,r)$ at a surface point against a dense ground-truth angular radiance sampled from the renderer; the paper's assumption predicts that the residual rises as light distance shrinks.","supporting_citations":[{"cited_title":"Structure-from-motion revisited","cited_arxiv_id":null,"evidence_quote":"Provides the sparse voxel grid implementation and fusion kernel ProbeSDF modifies, plus the MVMannequins dataset and the primary speed/accuracy baseline MMH."},{"cited_title":"Neural directional encoding for efficient and accurate view-dependent appearance modeling","cited_arxiv_id":null,"evidence_quote":"Provides the dense voxel SDF baseline Voxurf and the Gaussian smoothing regularization on raw SDF values that ProbeSDF adapts."},{"cited_title":"Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction","cited_arxiv_id":null,"evidence_quote":"Is the hash-grid baseline NeuS2 and the comparison point for training speed and grid-aligned artifacts."},{"cited_title":"2d gaussian splatting for geometrically accu- rate radiance fields","cited_arxiv_id":null,"evidence_quote":"Is the DTU baseline 2DGS whose no-mask protocol and half-resolution evaluation ProbeSDF matches for comparison."},{"cited_title":"Millimetric Human Surface Capture in Minutes","cited_arxiv_id":null,"evidence_quote":"Motivates conditioning appearance on the reflected vector rather than the view vector, which ProbeSDF adopts for probe sampling."}],"review_version":1}