{"id":"224804da-ca2a-47a1-8970-04f6f96e8c59","arxiv_id":"2412.05179","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A learned spatial mask modulates multi-resolution hash grid features per scene location, improving neural surface reconstruction accuracy on DTU and Tanks and Temples.","lead":"This paper teaches a neural network to choose how detailed its 3D encoding should be at each point in a scene, instead of using one global setting. The result is cleaner and more accurate 3D surface reconstructions from photographs on two standard benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not yet attributable to the spatial mask: the re-run Neuralangelo baseline is unverified and no mask-disabled control exists; a fair three-way comparison is needed.","rationale":"The reader's weakest assumption correctly identifies the fidelity of the re-run Neuralangelo baseline as load-bearing. I agree that this is the main vulnerability: the paper reports a large gap between the published Neuralangelo (0.61 mm) and its re-run (0.96 mm), yet gives no training-budget details that would let a reader determine whether the comparison is fair. I would extend the concern one step further: even if the re-run were faithful, the paper never ablate the mask itself within its own pipeline, so the improvement cannot be cleanly attributed to the proposed mechanism rather than to the other modifications and training choices. The mask mechanism is clearly described and the qualitative heatmaps are suggestive, but they do not substitute for a controlled comparison. Table 3 also shows the chosen mask-grid configuration (5-11, 0.654 mm) is not the best in the paper's own sweep (4-10, 0.648 mm), and with single runs and no error bars, the numerical advantage over published Neuralangelo is fragile. These are addressable with one controlled experiment; the conditional verdict remains appropriate. I therefore keep the verdict unchanged but stress that the proposed three-way test is the minimum evidence needed to support the state-of-the-art claim.","tokens_in":9565,"tokens_out":5987,"duration_ms":66349,"concrete_test":"Run the proposed pipeline on the 15 DTU scenes under three conditions with identical seeds, iterations, and loss weights: (A) official Neuralangelo unmodified, (B) the proposed pipeline with the mask fixed to s_l(x)=1 for all levels and positions (mask disabled but all other components present), and (C) the learned mask as in the paper. If the mean Chamfer-L1 obeys C < B and B approximately equals A by at least 0.1 mm consistently across scenes, the mask attribution holds. If B approximately equals C or B < C, the method's gain is not from spatial masking. Additionally, verify whether condition A reproduces the published 0.61 mm rather than 0.96 mm under the stated training budget; if it gives 0.96 mm, the baseline protocol differs from Neuralangelo's original setup and the comparison must be recalibrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the spatial mask causes the accuracy gain. The evidence is Tables 1-2, where the comparison baseline is either the published Neuralangelo numbers (0.61 mm DTU, 0.50 F1 TNT) or a re-run of the official repo (0.96 mm DTU, 0.47 F1 TNT). The re-run is far worse than published, but the paper does not state the iteration count, learning-rate schedule, or other training settings for either the re-run or the proposed method. If the re-run used a shorter or different budget than Neuralangelo's published protocol, part or all of the reported gain could be training budget rather than spatial masking. Critically, no ablation disables the mask inside the proposed pipeline by fixing s_l(x)=1, so the gain cannot be attributed to the mask instead of the progressive-unveiling schedule, gradient blocking, or the extra mask-network parameters. The reported margins over published Neuralangelo are small (0.05 mm DTU, 0.03 F1 TNT) and every number is a single run with no error bars. These issues together mean the state-of-the-art claim is plausible but not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a spatially-adaptive variant of multi-resolution hash encoding for neural surface reconstruction. Building on Neuralangelo, it learns a per-level scalar mask field, computed from a small hash grid and a shallow MLP, that multiplies the features of each hash-grid level as a function of spatial position. Training uses an RGB rendering loss, eikonal and curvature regularization, and a progressive unveiling of high-resolution grids with gradient blocking for inactive masks. Experiments on DTU and Tanks and Temples report mean Chamfer-L1 of 0.56 mm and F1 of 0.53, which the paper states is state-of-the-art compared with published Neuralangelo numbers of 0.61 mm and 0.50. Ablations examine mask grid resolution, curvature regularization, and sigmoid vs. softmax activations.","tokens_in":9775,"tokens_out":6143,"duration_ms":55957,"significance":"If the claimed improvement is real, the idea is simple, inexpensive, and likely to transfer to other grid-based neural fields. The paper makes a clear empirical contribution: the mask heatmaps provide a useful diagnostic, and the ablations of design choices are appropriate. However, the central SOTA claim depends on a comparison to a re-run Neuralangelo baseline whose training protocol is not documented, on small margins over published numbers, and on single-run results with no variance. No code or data are provided for verification. These gaps mean the result is plausible but not yet established.","major_comments":[{"comment":"The re-run of the official Neuralangelo repository (Neuralangelo †) reports a DTU mean Chamfer-L1 of 0.96 mm, compared with the published 0.61 mm. The manuscript does not state the number of training iterations, learning-rate schedule, loss weights, or other hyperparameters for this re-run or for the proposed method. If the re-run used a different or shorter training budget, the 0.40 mm difference between the re-run and the proposed method could be due to training budget rather than the spatial mask. Please provide the exact training configuration and, ideally, reproduce the published Neuralangelo numbers under the same protocol, and report the proposed method under the same protocol.","section":"Sec. 4.1, Table 2"},{"comment":"There is no ablation that disables the spatial mask within the proposed pipeline (e.g., fixing s_l(x)=1 for all levels). The progressive unveiling schedule, the gradient blocking to inactive masks, and the additional mask-network parameters are all confounded with the mask itself. A three-way comparison is needed: the full proposed method, the proposed method with the mask fixed to one, and the official Neuralangelo re-run under identical training settings. Without this, the reported gains cannot be attributed specifically to the spatially adaptive mask.","section":"Sec. 3.2 and Tables 1-2"},{"comment":"In the mask grid-resolution ablation, the default setting [5,11] yields Chamfer 0.654 mm, while [4,10] yields the best value of 0.648 mm. The text says the coarsest setting 'highlights the importance' of the mask but does not explain why [5,11] is preferred over [4,10]. Please justify the choice of default or update the main tables if the best configuration differs.","section":"Sec. 4.3, Table 3"},{"comment":"All results are from a single run with no error bars or significance testing. The mean margins over published Neuralangelo are small (0.05 mm Chamfer on DTU and 0.03 F1 on TNT), and scene-level differences (e.g., DTU Scan 69, where Ours gives 0.54 vs. published 0.53) are within typical run-to-run variation. Please report multiple seeds or a statistical analysis to support the claim that the differences are not noise.","section":"Tables 1-2"},{"comment":"The state-of-the-art claim is scoped to the methods listed in the tables, which are mostly from 2022-2023. Please either clarify that the claim is restricted to the compared baselines or add comparisons to more recent surface reconstruction methods that report DTU/TNT metrics. As written, the abstract's 'state-of-the-art performance on two benchmark datasets' is stronger than the evidence.","section":"Abstract and Tables 1-2"}],"minor_comments":[{"comment":"The sentence 'The results are detailed in Section 4.1' is incorrect; the Tanks and Temples results are in Section 4.2, while Section 4.1 presents DTU. Please fix the cross-reference.","section":"Sec. 4.2"},{"comment":"There are grammatical and typographical errors: 'offical' should be 'official', and 'on both DTU and as well as Tanks and Temples' should be 'on both DTU and Tanks and Temples'.","section":"Sec. 4.2 and Discussion"},{"comment":"The grid resolution range '323 to 20483' should read '32^3 to 2048^3' with proper superscripts.","section":"Sec. 4.1"},{"comment":"The caption should state the color scale and the exact level groupings (e.g., levels 1-8, 8-14, 15-16) used for the Low/Mid/High-Frequency heatmaps, rather than leaving this to the body text.","section":"Figure 4"},{"comment":"The paper states that code will be released after acceptance; for reproducibility, please provide code or a detailed configuration listing all hyperparameters at submission time.","section":"General"},{"comment":"Consider clarifying that the mask s_l(x) is a scalar multiplier shared across all feature channels of level l, since the notation s_l(x) * f_l could also be read as channel-wise multiplication if f_l is treated as a vector.","section":"Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the unverified re-run baseline and the absence of a mask-disabled ablation. If the authors can provide a faithful Neuralangelo baseline under identical training settings and a within-pipeline control, the paper could become acceptable. I would not reject on the current evidence, but the central claim is not yet established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is genuinely tidy: replace the fixed, global multi-resolution hash encoding in Neuralangelo with a learned per-level scalar mask that varies in space. The implementation is cheap (a small hash grid and a 16-unit MLP, about 2% extra parameters) and the qualitative mask visualizations are the best part of the paper—you can see the high-frequency grids switching on at edges and switching off on smooth surfaces, which is exactly the behavior you want.\n\nWhat the paper does well: the method is simple, clearly described, and the ablations cover the important design choices (mask grid resolution, activation function, curvature loss). The connection to SAPE is made honestly, and the limitations section is refreshingly direct about memory cost and reflective scenes. The idea is a natural extension of SAPE from sinusoidal encodings to hash grids, and it is not present in prior work.\n\nThe soft spots are real and load-bearing. The headline SOTA claim rests on comparing against a re-run of Neuralangelo (labeled †) that scores 0.96 mm on DTU versus the published 0.61 mm. The paper gives no training settings for that re-run, so the 0.40 mm gap in the baseline could come from a shorter budget or different hyperparameters, not from the spatial mask. And there is no ablation that disables the mask inside the proposed pipeline by fixing s_l(x)=1, so even the within-pipeline comparison cannot isolate the mask's contribution. The margins over the published Neuralangelo numbers are small (0.05 mm DTU, 0.03 F1 TNT), every number is a single run with no error bars, and the chosen mask resolution [5,11] is not the best in the paper's own ablation (0.654 mm vs. 0.648 mm for [4,10]). No code is released. These are all addressable, but they mean the central claim is plausible, not established.\n\nWho this is for: anyone working on neural surface reconstruction, especially on grid-based SDF methods. The mechanism is simple enough that a reader could reimplement it in an afternoon, and the qualitative evidence suggests the idea works even if the exact numbers are shaky.\n\nRecommendation: this deserves a serious referee, not a desk reject. The concept is sound and the writing is clear; the weaknesses are experimental. A referee should ask for a three-way comparison—published Neuralangelo, a re-run with matching training settings, and the proposed method with the mask disabled—plus a few repeats to get error bars. That would turn a plausible contribution into a solid one.","headline":"A clean, plausible extension of Neuralangelo with a per-level learned spatial mask over hash features, but the SOTA claim is not yet attributable to the mask: the re-run baseline is unverified and no mask-disabled control exists.","tokens_in":10344,"tokens_out":1912,"would_cite":false,"duration_ms":20474,"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":"A learned spatial mask over multi-resolution hash-grid features lets neural surface reconstruction use fine grids where detail exists and coarse grids elsewhere, achieving state-of-the-art Chamfer-L1 of 0.56 mm on DTU and F1 of 0.53 on…","keywords":["neural surface reconstruction","hash encoding","spatially adaptive encoding","signed distance function","volume rendering","multi-resolution hash grid","DTU dataset","Tanks and Temples"],"falsifier":"Run the proposed method and the official Neuralangelo baseline from identical code, seeds, iteration counts, and hyperparameters, and check whether the mean DTU Chamfer-L1 gap stays near the reported 0.40 mm; if the gap shrinks or the method no longer beats the published 0.61 mm, the spatial-mask explanation for the gain fails. A complementary check is to inspect the learned high-frequency mask over a known flat region in a synthetic scene, where it should be near zero.","tokens_in":9273,"feed_emoji":"🗺️","tokens_out":8419,"duration_ms":78869,"temperature":0.7,"pith_summary":"Surface reconstruction from images with neural signed distance functions usually relies on a fixed stack of hash grids, which biases the network toward either coarse or fine geometry everywhere at once. This paper argues that the right bias is spatial: smooth regions should be encoded with coarse grids, while edges and fine texture should use high-resolution grids. It introduces a learned mask field that multiplies the features of each hash-grid level by a per-location weight, so the network effectively chooses its encoding basis as a function of position, together with a progressive schedule that unveils finer grids over training. On the DTU and Tanks and Temples benchmarks, the method reports state-of-the-art accuracy, including a mean Chamfer-L1 of 0.56 mm on DTU against 0.96 mm for the re-run Neuralangelo baseline.","feed_headline":"Space-aware hash grids cut neural surface error to 0.56 mm","feed_subtitle":"A learned mask lets each region pick its hash-grid resolution, beating Neuralangelo on DTU and Tanks and Temples.","key_machinery":"The load-bearing machinery is the learned spatial mask field. A small hash grid (8 levels, feature dimension 4, hash size $2^{18}$) plus a one-hidden-layer MLP with sigmoid activation outputs $N$ scalar weights $s_\\ell(x)$, one per level $\\ell$ of the main 16-level SDF hash grid (resolutions $32^3$ to $2048^3$, feature dimension 8, hash size $2^{22}$). The final encoding multiplies each level's feature vector by its mask weight before concatenation, so the network can silence any resolution in any region. Progressive unveiling of fine grids, and gradient blocking to masks of inactive grids, prevents the mask from permanently suppressing high frequencies early in training. Numerical-gradient $\\epsilon$ scheduling is inherited from the backbone to provide coarse-to-fine surface optimization.","core_discovery":"The central discovery is that a spatially varying mask over hash-grid resolutions is enough to fit a wider range of frequencies without adding noise. The mask is produced by a small auxiliary hash grid and a shallow sigmoid MLP, yielding per-level scalars $s_\\ell(x) \\in (0,1)$; the SDF hash encoding becomes $h(x) = [s_1(x) f_1, \\dots, s_N(x) f_N]$, and progressive unveiling blocks gradients to inactive mask levels early in training. Learned masks segment scenes by frequency need: fine grids activate on edges and gratings, while smooth or specular regions fall back to coarse grids. The paper shows this leads to cleaner low-frequency surfaces and improved fine details, with quantitative gains over Neuralangelo on both DTU and Tanks and Temples.","pith_inferences":["If the learned masks genuinely reflect frequency need, they could be reused after training as a pruning map: hash entries in regions where all fine-level masks are near zero are never read, so the about 4.2 GB footprint could be cut without retraining. The paper gestures at this as future work.","A controlled test of the mechanism would force the high-frequency masks to be active in a smooth region and measure the resulting chamfer error; the heat maps in Figure 4 suggest this should hurt, but the paper does not run that experiment.","The spatial-mask idea is orthogonal to the choice of encoding beyond hash grids; coupling it with spatially adaptive view-direction encodings could address reflective scenes, a limitation the paper acknowledges. This is a natural next step the authors name."],"forward_implications":["On the 15-scene DTU benchmark, the method reports a mean Chamfer-L1 of 0.56 mm, below both the re-run Neuralangelo (0.96 mm) and the published Neuralangelo number (0.61 mm).","On Tanks and Temples, the method's F1 score is 0.53, above the re-run (0.47) and published (0.50) Neuralangelo numbers, with visible gains on structures like the barn grating and truck wheels.","Ablations show the sigmoid mask activation is important: replacing it with softmax drops performance (Chamfer 0.74 vs 0.65) because the network collapses toward the coarsest resolution.","The mask network adds only about 8 million parameters (374M vs 366M), so the adaptive bias comes at negligible storage cost beyond the hash-grid features themselves."],"supporting_citations":[{"why":"Provides the Neuralangelo backbone: multi-resolution hash-grid SDF, numerical-gradient epsilon scheduling, and the official-code baseline the method extends and is compared against.","marker":"[9]"},{"why":"Introduces spatially adaptive progressive encoding (SAPE), the direct inspiration for per-region frequency selection and progressive unveiling of fine resolutions.","marker":"[7]"},{"why":"Defines the multi-resolution hash encoding that the proposed spatial masks modulate.","marker":"[13]"},{"why":"Establishes NeuS volume rendering of neural SDFs, the rendering and surface-extraction framework used at test time.","marker":"[22]"},{"why":"Supplies the DTU benchmark with structured-light ground truth point clouds and the Chamfer-L1 evaluation protocol.","marker":"[1]"},{"why":"Supplies the Tanks and Temples benchmark with laser-scanned ground truth and the F1 evaluation script.","marker":"[8]"}],"fun_headline_variants":["Adaptive hash grids let regions pick resolution, beat Neuralangelo","Spatially masked hash encodings sharpen neural surface reconstruction","Learned spatial masks over hash grids give cleaner surfaces","Per-location hash resolution improves neural surface fitting","Hash encodings that adapt in space achieve state of the art"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported state-of-the-art result rests on the authors' re-run of the official Neuralangelo code, which scores 0.96 mm on DTU rather than the 0.61 mm in the original Neuralangelo paper; if that re-run is not a faithful like-for-like baseline, the improvement attributed to the spatial mask is not established.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive hash grids let regions pick resolution, beat Neuralangelo","Spatially masked hash encodings sharpen neural surface reconstruction","Learned spatial masks over hash grids give cleaner surfaces","Per-location hash resolution improves neural surface fitting","Hash encodings that adapt in space achieve state of the art"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000436,"raw_usage":{"total_tokens":2171,"prompt_tokens":851,"completion_tokens":1320,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":1238}},"tokens_in":467,"tokens_out":1320,"duration_ms":11095,"temperature":1.0,"reasoning_tokens":1238,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:50:28.807840+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed method and the official Neuralangelo baseline from identical code, seeds, iteration counts, and hyperparameters, and check whether the mean DTU Chamfer-L1 gap stays near the reported 0.40 mm; if the gap shrinks or the method no longer beats the published 0.61 mm, the spatial-mask explanation for the gain fails. A complementary check is to inspect the learned high-frequency mask over a known flat region in a synthetic scene, where it should be near zero.","supporting_citations":[{"cited_title":"Neuralangelo: High-fidelity neural surface reconstruction","cited_arxiv_id":null,"evidence_quote":"Provides the Neuralangelo backbone: multi-resolution hash-grid SDF, numerical-gradient epsilon scheduling, and the official-code baseline the method extends and is compared against."},{"cited_title":"Sape: Spatially-adaptive progressive encoding for neural optimization","cited_arxiv_id":null,"evidence_quote":"Introduces spatially adaptive progressive encoding (SAPE), the direct inspiration for per-region frequency selection and progressive unveiling of fine resolutions."},{"cited_title":"Instant neural graphics primitives with a multires- olution hash encoding","cited_arxiv_id":null,"evidence_quote":"Defines the multi-resolution hash encoding that the proposed spatial masks modulate."},{"cited_title":"Large-scale data for multiple-view stereopsis","cited_arxiv_id":null,"evidence_quote":"Supplies the DTU benchmark with structured-light ground truth point clouds and the Chamfer-L1 evaluation protocol."},{"cited_title":"Tanks and temples: Benchmarking large-scale scene reconstruction","cited_arxiv_id":null,"evidence_quote":"Supplies the Tanks and Temples benchmark with laser-scanned ground truth and the F1 evaluation script."}],"review_version":1}