{"id":"04e598b0-07d1-47cb-908d-7b491181d931","arxiv_id":"2501.01196","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Inter-image feature matching, not monocular depth, supplies the depth prior that enables accurate neural implicit surface reconstruction from sparse indoor views.","lead":"This paper introduces Sparis, a method that reconstructs 3D indoor surfaces from as few as 10 to 20 photos by using matching information between image pairs instead of unreliable monocular depth estimates. The result is a large jump in reconstruction accuracy on standard benchmarks, which matters for augmented reality, robotics, and 3D scanning where capturing hundreds of photos is impractical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core claim that inter-image matching yields accurate absolute depth priors for sparse indoor reconstruction is not supported by any direct measurement of RoMa match density or triangulated-depth error in the evaluation scenes; without such measurement, the reported F-score gains cannot be…","rationale":"The reader's weakest assumption matches my own: the method relies on a pretrained dense matching network to provide geometrically consistent correspondences under sparse indoor views, yet the paper provides no direct evidence about match density, matching accuracy, or triangulated-depth quality. That is the single most load-bearing concern because the inter-image depth and reprojection losses are the only new supervision signals that distinguish Sparis from prior prior-based neural implicit methods. If RoMa matches are sparse or inaccurate, especially on the textureless surfaces that dominate indoor scenes, the central claim that the method provides 'more accurate depth information' collapses, even though the overall pipeline may still work reasonably through the normal prior and the neural rendering regularization. I am not arguing that the method is fraudulent or that its results are impossible; rather, the current evaluation underdetermines the causal contribution of the proposed inter-image prior. A direct preprocessing-stage diagnostic would settle the point quickly and is a standard verification step. Since the reader already issued a CONDITIONAL verdict largely for this reason, I see no need to change the verdict; I would keep it CONDITIONAL pending the match-quality measurement and, ideally, released code and per-scene numbers with baselines run under identical conditions.","tokens_in":19747,"tokens_out":4196,"duration_ms":46223,"concrete_test":"Reproduce the preprocessing stage on the same 10 ScanNet and 8 Replica scenes: run RoMa on all ordered image pairs, apply the angular filter (Eq. 10-11) and the epipolar weighting (Eq. 12), then for each reference view compute (a) the percentage of image pixels with surviving matches, (b) the median Sampson distance, and (c) the absolute error of triangulated depth eD against ground-truth depth for those pixels. Plot per-scene match coverage and depth error against the per-scene F-score improvement over the no-depth baseline. If median triangulated-depth error exceeds a few centimeters, or coverage drops below a small fraction of pixels on planar untextured regions, the 'more accurate depth information' claim is not substantiated and the method should be re-evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that RoMa matches, after the confidence and epipolar filters, are dense and accurate enough on textureless indoor surfaces to provide reliable absolute depth anchors. The method's only data-dependent priors are the triangulated depths eD in Eq. (7)/(14) and the reprojection targets in Eq. (9)/(15), both derived from Eq. (6). The paper never reports how many matches survive per view pair, how their Sampson distances are distributed, or how far triangulated eD is from ground-truth depth on the actual ScanNet/Replica scenes. On walls, floors, and ceilings, where monocular priors also struggle, feature matching can be sparse or dominated by low-confidence estimates; the angular filter (Eq. 10-11) only discards entire source views, and the epipolar weight (Eq. 12) only down-weights matches, it cannot invent correspondences where none exist. Therefore the abstract's claim that the inter-image prior 'offers more accurate depth information' is not directly verified. The large F-score gaps (0.647 vs 0.464 on ScanNet; 0.825 vs 0.454 on Replica) could in principle stem from the retuned baselines, the exclusion of failed scenes, or the normal prior, rather than from the inter-image depth anchors, leaving the central causal claim underdetermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Sparis, a neural implicit surface reconstruction method for indoor scenes from sparse views. Instead of relying on monocular depth priors, it uses inter-image feature matching: RoMa correspondences between view pairs are triangulated into absolute depth targets, which supervise a VolSDF-based neural surface rendering; a reprojection loss encourages cross-view consistency; and an angular filter plus an epipolar weight suppress unreliable matches. The method is evaluated on ScanNet (15-20 views) and Replica (10 views), where it reports large F-score improvements over NeuRIS, MonoSDF, and other baselines, and ablations show each component contributes. The supplementary provides per-scene tables, additional comparisons with Gaussian-splatting methods and DUSt3R, and a discussion of pose assumptions.","tokens_in":20136,"tokens_out":4963,"duration_ms":49572,"significance":"If the results are reliable, Sparis offers a meaningful step for sparse-view indoor reconstruction by replacing scale-ambiguous monocular depth priors with absolute depths triangulated from learned correspondences. The idea is simple and uses an off-the-shelf matcher, and the paper includes useful ablations and per-scene supplementary tables. The central claim, however, is weakened by evaluation choices: a key baseline is retuned from its default, failed scenes are excluded from averages, no error bars or seed variance are reported, and the claimed accuracy of the inter-image depth prior is never directly measured. The method is not conceptually circular, but the empirical support for its main causal claim is incomplete.","major_comments":[{"comment":"The baseline comparisons are not even-handed. MonoSDF is changed from its default monocular depth weight (0.1) to 0.001 because the default 'unable to produce valid meshes,' and NeuS and HelixSurf averages are computed only over scenes where they produced valid meshes (4 and 1 failures, respectively, out of 10 ScanNet scenes). This selective reporting biases the headline improvements (0.647 vs 0.464 on ScanNet, 0.825 vs 0.454 on Replica). Please report per-scene results for all methods, count failed scenes as F-score 0 or as a separate 'failed' category, and provide default-vs-retuned MonoSDF numbers so readers can assess the effect of the retuning.","section":"Comparison, Tables 1-2"},{"comment":"The central claim that inter-image matching provides 'more accurate depth information' is not directly tested. The paper never reports the number of RoMa matches per view pair, the distribution of Sampson distances, or the error of the triangulated depths eD(r) relative to ground-truth depth on the evaluation scenes. On textureless walls, floors, and ceilings, match density could be low and the angular filter and epipolar weight cannot invent missing correspondences. Without diagnostics on match density and triangulation accuracy, the large F-score gains cannot be attributed to the proposed depth prior rather than to the normal prior or to the reprojection loss acting as regularization. Please add such quantitative analysis for ScanNet and Replica.","section":"Inter-Image Depth Loss, Eqs. (7)-(15)"},{"comment":"The main comparisons in Tables 1 and 2 use ground-truth poses for all methods, while COLMAP poses are evaluated only for the proposed method (Table 5). Since the triangulated depth prior is directly sensitive to pose accuracy, the comparison is not symmetric. Please run MonoSDF, NeuRIS, HelixSurf, and the proposed method with the same COLMAP poses and report the results; if those baselines are too pose-sensitive to run, state that limitation explicitly and justify why the GT-pose comparison is the relevant one.","section":"Experiments, Table 5 and Section D of Supplementary"},{"comment":"No measure of variability is reported. With only 10 ScanNet scenes and 8 Replica scenes, and with baselines retuned or selectively averaged, the reported gaps could be within training stochasticity. Please report mean and standard deviation over multiple seeds for the proposed method and the main baselines, or otherwise justify that the results are stable. This is especially important because the neural rendering training and the sampling-based losses are stochastic.","section":"Ablation Study and Table 3; Tables 1-2"}],"minor_comments":[{"comment":"The term '1ui r,s' appears to be a typo; it should be '(1 − ui r,s)' as in Eq. (7).","section":"Eq. (14)"},{"comment":"'Anaysis' is a typo and should be 'Analysis'.","section":"Section 'Experiments and Anaysis'"},{"comment":"The list of ablation settings repeats the numbering '(4)' twice; the settings should be numbered 1 through 5.","section":"Ablation Study"},{"comment":"The footnote for DUSt3R says 'GT poses are included as inputs,' but DUSt3R is designed to operate without poses; please clarify what was provided and why this is a fair comparison.","section":"Supplementary Table 4"},{"comment":"For reproducibility, please specify the exact number of views used for each ScanNet scene, the camera intrinsics used, and whether a code release is planned.","section":"Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the method appears technically sound, but the evaluation needs substantial strengthening before the central claims can be endorsed. The selective baseline handling and the absence of pose-robustness and match-quality diagnostics are the main concerns. I would also encourage the editor to ask for a direct comparison of the triangulated depth prior against monocular depth priors on the same scenes, since that is the paper's stated motivation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Wu et al.'s Sparis is a clean, incremental extension: replace scale-ambiguous monocular depth with triangulated depth from RoMa feature matches, add a reprojection loss and two filters, and otherwise keep the VolSDF/MonoSDF skeleton. That's a genuine new combination, and the reported gains on ScanNet (0.647 vs 0.464 F-score) and Replica are large if they hold. The ablations in Table 3 are the most useful part: each component (normal prior, inter-image depth, reprojection) contributes, and the supplementary comparison with scaled monocular depth (Lscaledepth) gives a fairer test of the paper's central claim than the headline tables.\n\nThe soft spots are all in the evaluation. MonoSDF is retuned from its published configuration by lowering the depth weight from 0.1 to 0.001, which amounts to neutering its main prior; NeuS and HelixSurf are averaged only over scenes where they produced valid meshes; and there are no error bars or seed variance. The supplementary per-scene tables help, but readers shouldn't have to re-aggregate to check which scenes were excluded. The core causal claim — that inter-image matching gives more accurate depth — is also not directly measured anywhere. The paper never reports RoMa match density, Sampson-distance distributions, or triangulated-depth error on the test scenes. That said, I don't think the stress-test note's strong version lands: the ablation of Ldepth does isolate its contribution, so the claim isn't circular or unsupported; it's just under-measured. A revision could settle it with one plot of eD against GT depth.\n\nCitation pattern is fine; the only self-citation is NeuSurf, used as a baseline. Who is this for: practitioners in sparse-view indoor reconstruction and anyone extending prior-based implicit surfaces. It deserves a serious referee; the method is coherent and the engineering is solid. I'd send it to peer review, with the expectation that the authors tighten the baseline tuning and add a direct validation of the depth prior.","headline":"Solid engineering extension of VolSDF with matching-based depth priors; the headline gains are real-looking but the evaluation needs tightening and the depth-prior claim is under-measured.","tokens_in":20589,"tokens_out":2938,"would_cite":false,"duration_ms":29358,"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":"Sparis reconstructs indoor surfaces from 10–20 images by replacing monocular depth with triangulated inter-image matches.","keywords":["sparse-view reconstruction","indoor scene reconstruction","neural implicit surface","signed distance function","inter-image matching","depth prior","cross-view reprojection","3D reconstruction"],"falsifier":"Take a sparse indoor sequence with large textureless regions and wide-baseline image pairs, run the matching network, and compare the density and accuracy of the triangulated depth anchors against ground-truth depth. If match coverage is too low or the Sampson-weighted correspondences still drift, the reported F-score advantage over monocular-prior baselines would shrink or disappear. A controlled ablation that swaps in a weaker matching network while keeping all other training fixed should degrade reconstruction quality in proportion to match quality.","tokens_in":19512,"feed_emoji":"🏠","tokens_out":6900,"duration_ms":59151,"temperature":0.7,"pith_summary":"The paper claims that sparse-view indoor reconstruction collapses because monocular depth priors cannot be scaled correctly from only a few views. It replaces monocular depth with an inter-image matching prior: matched pixel pairs across views are triangulated with known camera poses to give absolute depth anchors, and a reprojection loss ties the rendered surface back to the matched pixels. Two filters, an angular score for choosing source views and an epipolar weight for downweighting geometrically inconsistent matches, make the prior robust to matching errors. On ScanNet with 15–20 views the method reaches an F-score of 0.647 versus 0.464 for NeuRIS, and on Replica with 10 views 0.825 versus 0.454 for MonoSDF. If correct, practical room-scale 3D reconstruction from a handful of photos becomes feasible with neural implicit surfaces.","feed_headline":"Triangulated pixel matches rebuild indoor 3D from 10 photos","feed_subtitle":"Sparse views no longer need monocular depth scaling: inter-image triangulation lifts ScanNet F-score to 0.647 from 0.464.","key_machinery":"The load-bearing object is the inter-image matching prior built from a pretrained dense feature matching network. For each image pair the network returns pixel correspondences and a confidence; with known camera poses these are triangulated into absolute depth values that supervise the neural SDF renderer. The two supporting mechanisms are an angular filter, a certainty-weighted angular score used to select the source view with favorable triangulation geometry, and an epipolar weight function that downweights matches by their Sampson distance to the epipolar constraint, as expressed in equations (12)–(15). Together they convert raw matching output into depth and reprojection supervision that is less sensitive to matching noise.","core_discovery":"The central claim is that inter-image correspondence, not monocular depth, is the right geometric prior for sparse-view indoor surface reconstruction. The authors argue that monocular depth supervision requires estimating a global scale and shift, which is ill-posed when overlap between views is small, so the geometry collapses. Instead they extract dense pixel matches between image pairs, triangulate those matches into 3D points and hence absolute depth, and supervise the signed-distance-field renderer with an inter-image depth loss. A cross-view reprojection loss then forces the rendered surface point along a ray to reproject onto the matched pixel in the other view, enforcing consistency across views. The angular filter and epipolar weight reduce the influence of wrong or weakly constrained matches. With these pieces, the method reportedly produces smoother, more complete and more accurate meshes than prior indoor reconstruction methods under the same sparse-view settings.","pith_inferences":["A natural extension the paper leaves implicit is refining the matching network or fusing it into the training loop, since improvements in match quality should directly improve triangulated depth and reprojection.","The COLMAP-pose experiment implies pose error costs roughly 0.13 F-score; joint refinement of poses and geometry could recover part of that gap.","Scenes with repetitive texture or wide baselines will stress the pairwise matcher, so the practical operating envelope is set by match density on textureless surfaces, which the paper does not quantify.","The inter-image depth and reprojection scheme could be inserted into other neural field or Gaussian-splatting renderers, not only the SDF-based pipeline used here."],"forward_implications":["Indoor surface reconstruction no longer requires hundreds of views: with 10–20 images the method reports F-scores of 0.647 on ScanNet and 0.825 on Replica, ahead of the monocular-prior baselines it compares against.","Monocular depth, even when scaled optimally by least squares, is the point of failure in the sparse regime; replacing it with absolute triangulated depth removes the scale ambiguity.","Cross-view reprojection consistency acts as a regularizer that reduces overfitting when view overlap is low.","The angular filter and epipolar weight make the reconstruction resilient to matching noise, and this robustness carries over to estimated camera poses: with COLMAP poses the method still reports 0.514 F-score, above 0.464 for NeuRIS with ground-truth poses.","The same priors transfer to object-level sparse reconstruction, giving Chamfer distance comparable to the leading object-level method on DTU with 3 views."],"supporting_citations":[{"why":"Supplies the pretrained dense feature matching network whose pixel correspondences and confidences are the source of the inter-image depth and reprojection priors.","marker":"Edstedt et al. 2023"},{"why":"VolSDF is the neural implicit surface rendering baseline whose SDF-volume rendering pipeline Sparis adopts.","marker":"Yariv et al. 2021"},{"why":"NeuRIS provides the monocular normal prior component and is the strongest comparison baseline on both datasets.","marker":"Wang et al. 2022a"},{"why":"MonoSDF is the monocular-depth-prior method whose scale/shift depth fitting Sparis identifies as the failure mode in sparse views.","marker":"Yu et al. 2022"},{"why":"Omnidata supplies the pretrained normal predictions used as the normal prior in the total loss.","marker":"Eftekhar et al. 2021"},{"why":"DDP-NeRF defines the sparse-view ScanNet sampling setting (15–20 images) and is one of the comparison baselines.","marker":"Roessle et al. 2022"},{"why":"Supplies the Eikonal regularization term on the SDF field used in the overall loss.","marker":"Gropp et al. 2020"}],"fun_headline_variants":["Pixel matches triangulate sparse views into indoor 3D","Inter-image triangulation rebuilds indoor 3D from few photos","Sparse views get accurate depth via cross-view triangulation","Triangulation lifts sparse-view indoor reconstruction","Pixel correspondences replace scale-ambiguous depth in sparse 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the pretrained feature matching network supplies enough correct, dense, geometrically consistent correspondences across sparse indoor views, especially on textureless walls and ceilings and over wide baselines; if matches are too sparse or noisy, triangulated depth anchors and reprojection constraints lose accuracy and the two filters cannot repair the missing matches. The main comparisons also assume known camera poses.","fun_headline_variants_meta":{"raw":{"variants":["Pixel matches triangulate sparse views into indoor 3D","Inter-image triangulation rebuilds indoor 3D from few photos","Sparse views get accurate depth via cross-view triangulation","Triangulation lifts sparse-view indoor reconstruction","Pixel correspondences replace scale-ambiguous depth in sparse 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000506,"raw_usage":{"total_tokens":2449,"prompt_tokens":905,"completion_tokens":1544,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":1462}},"tokens_in":521,"tokens_out":1544,"duration_ms":11661,"temperature":1.0,"reasoning_tokens":1462,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:32:23.857303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sparse indoor sequence with large textureless regions and wide-baseline image pairs, run the matching network, and compare the density and accuracy of the triangulated depth anchors against ground-truth depth. If match coverage is too low or the Sampson-weighted correspondences still drift, the reported F-score advantage over monocular-prior baselines would shrink or disappear. A controlled ablation that swaps in a weaker matching network while keeping all other training fixed should degrade reconstruction quality in proportion to match quality.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Omnidata supplies the pretrained normal predictions used as the normal prior in the total loss."},{"cited_title":"T.; Mildenhall, B.; Srinivasan, P","cited_arxiv_id":null,"evidence_quote":"DDP-NeRF defines the sparse-view ScanNet sampling setting (15–20 images) and is one of the comparison baselines."}],"review_version":1}