{"id":"c926104b-62dd-45f0-b70c-76d70bb74e5a","arxiv_id":"2504.20378","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An MVS-initialized 2D Gaussian Splatting method with fixed-color feature splatting, reparameterized disk sampling, and selective Gaussian updates improves sparse-view surface reconstruction on DTU to a Chamfer distance of 1.13.","lead":"Sparse2DGS is a Gaussian Splatting pipeline that initializes from multi-view stereo point clouds and adds geometry-prioritized losses to reconstruct 3D surfaces from only three input views. It reports lower Chamfer error on DTU scenes than prior Gaussian methods while training in about ten minutes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DGPR (Eq. 6) assumes CLMVSNet features are view-consistent at off-surface disk samples; in occluded or textureless regions this can yield spurious gradients, so the reported 1.13 CD gain from DGPR may rest on an unvalidated assumption.","rationale":"The reader's weakest assumption correctly identifies the DGPR feature-consistency requirement as the least secure link. My reading agrees: the paper's ablations attribute a small but real gain to DGPR, but the loss is applied without any visibility weighting or validation of feature consistency at off-surface points. This is not an ad hominem or a disagreement with consensus; it is a specific, testable assumption in Eq. 6. If the proposed occlusion-masked ablation shows a material change, the paper's central mechanism would need to be re-explained or the loss re-designed, so the CONDITIONAL verdict remains appropriate. I do not see a reason to move to ACCEPT or REJECT based on the current evidence. The absence of error bars, noted by the reader, is a secondary issue. One possible counter-argument: even if DGPR were removed, the method still achieves CD 1.201 with MVS init, fixed color and feature splatting, which beats PGSR (2.08). So the headline 'outperforms Gaussian methods' is robust. However, the paper's key insight and the claim of 'geometry-prioritized enhancement schemes' depend on DGPR being beneficial for the right reason. The proposed test would settle that. Therefore, I keep the verdict UNCHANGED (conditional acceptance with a request for this experiment).","tokens_in":13100,"tokens_out":12627,"duration_ms":130568,"concrete_test":"Ablate the DGPR loss on the 15 DTU scenes by adding an occlusion mask: only include disk samples whose projections are mutually visible, using rendered depth from the current Gaussians (or ground-truth depth) to determine visibility in both views. Compare the resulting CD to the reported 1.13; a change larger than ~0.05 would show that the unmasked L_df is not contributing the claimed geometric signal. As a second, complementary check, replace the CLMVSNet features used in Eq. 6 with random fixed features and keep all other losses; if DGPR still improves over no-DGPR, the view-consistency assumption is not load-bearing, whereas a drop to baseline would confirm it is.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the geometry-prioritized losses, especially the reparameterization-based disk sampling loss L_df (Eq. 6), turn MVS-initialized 2D Gaussians into accurate surfaces. Eq. 6 projects K points sampled on each Gaussian disk (Z = p + RS Z', Eq. 7) into a source and a target view, and compares CLMVSNet FPN features at the projected positions via cosine similarity. This is only a valid geometric signal if the feature maps are locally view-consistent at arbitrary off-surface points. The disks in 2DGS are local surface approximations, and the sampled points intentionally scatter around the disk; many will project into occluded, textureless, or repetitive areas in the target view. The paper never evaluates whether the fixed CLMVSNet features satisfy this consistency, nor does it mask occluded samples. The paper's own limitation section concedes that MVS errors under occlusion propagate to the method. If the feature consistency is unreliable, L_df can push Gaussians off the true surface to match spurious feature similarities, and the modest improvement from DGPR (CD 1.201 to 1.142 in Table 2) could be an artifact of this bias rather than evidence of correct geometric regularization. Because the title and abstract locate the key insight in these enhancement schemes, this unvalidated assumption is load-bearing for the paper's contribution, even though the MVS initialization alone already accounts for most of the gap to prior Gaussian baselines.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Sparse2DGS, a sparse-view surface reconstruction method built on 2D Gaussian Splatting (2DGS). It initializes the Gaussian cloud from the dense point cloud of the unsupervised MVS network CLMVSNet, then applies three geometry-prioritized enhancements: (i) fixed-color and fixed-feature supervision using MVS feature maps rendered through the Gaussian splatting pipeline (Sec. 3.2); (ii) a Direct Gaussian Primitive Regularization (DGPR) loss that reparameterizes each disk by sampling points from a standard Gaussian, transforming them by the primitive's position, rotation, and scale, and supervising the projected points with cross-view feature consistency (Sec. 3.3, Eqs. 6-8); and (iii) a Selective Gaussian Update (SGU) that compares patch-based NCC scores between warping using Gaussian primitives and warping using rendered depths/normals, then reprojects better rendered depths into updated positions (Sec. 3.4, Eq. 11). On the DTU dataset with three views (23, 24, 33), the method reports a mean Chamfer Distance of 1.13, outperforming 2DGS (2.81), GOF (2.82), and PGSR (2.08), while being competitive with NeRF-based sparse-view methods and about 2x faster than SparseNeuS. Ablations in Tables 2, 4, 5, and 6 attribute improvements to each proposed module.","tokens_in":13528,"tokens_out":5184,"duration_ms":54200,"significance":"If the results hold, this is a practical and timely contribution: it shows that a dense MVS point cloud combined with fixed-feature supervision and point-based disk regularization can make Gaussian splatting much more accurate for sparse-view surface reconstruction than standard SfM-initialized 2DGS, while keeping training time around ten minutes. The paper ships code, uses the official DTU evaluation protocol, and includes a controlled ablation that separates the MVS-initialization gain from the proposed losses. The main methodological novelty, DGPR, is clearly described and the cross-view consistency formulation is principled in the ideal case of perfectly view-consistent features. However, the central claim that the proposed regularization is responsible for the reported gains rests on an assumption about the local view consistency of fixed MVS features at off-surface samples, and the ablation margins are small and based on single runs, so the significance of the individual modules is not yet fully established.","major_comments":[{"comment":"The DGPR loss assumes that CLMVSNet FPN features are locally view-consistent at arbitrary points sampled on the 2D Gaussian disks, including points that do not lie on the true surface and that may project into occluded, textureless, or repetitive regions of the target view. The manuscript does not provide any validation of this assumption: there is no occlusion masking, no analysis of the fraction of off-surface samples that are actually visible in the target view, and no comparison against a baseline where the features are randomized or where only on-surface points are used. Since the paper's abstract and introduction attribute the key improvement to this geometric-prioritized regularization, this unvalidated assumption is load-bearing for the central claim. Please provide quantitative evidence of feature consistency, add a visibility mask or a robustness analysis, or ablate the loss against a random-feature version to show that the CD gain in Table 2 (row e to row f, 1.201 to 1.142) is not driven by spurious feature similarities.","section":"Tables 2, 5, and 6"},{"comment":"All experiments report a single run with no variance or seed information, and several ablation margins are small: Table 2 shows DGPR improving CD from 1.201 to 1.142 (0.059) and Table 5 shows SGU improving CD from 1.201 to 1.152 (0.049), while Table 6 shows the effect of K varying from 1.145 to 1.139. These differences are within the typical run-to-run variability of Gaussian splatting optimizations, so the per-module attribution is not statistically supported. Please report mean and standard deviation over at least three random seeds for the main ablations in Tables 2, 4, 5, and 6, and state whether the reported metrics are averaged or from a single representative run.","section":"Sec. 3.4, Eq. (11)"},{"comment":"The selective update rule replaces a Gaussian position whenever NCC_R > NCC_G, with no absolute threshold on the NCC values. In textureless or homogeneous regions, even a tiny positive difference between two low NCC scores would trigger an update, potentially injecting noise from unreliable rendered depth into the point set. The paper does not report the distribution of NCC scores, the fraction of primitives updated per iteration, or the sensitivity of the final CD to an NCC threshold. Please provide such an analysis (e.g., a sweep over thresholds, or a report of typical NCC values for updated versus non-updated primitives) to demonstrate that the update is meaningful and not dominated by low-confidence matches.","section":"Sec. 4.3, Table 1"}],"minor_comments":[{"comment":"Equation (12) contains a stray comma after \\lambda_2 L_n; please correct to \"\\mathcal{L} = \\mathcal{L}_{rgb} + \\lambda_1 \\mathcal{L}_d + \\lambda_2 \\mathcal{L}_n + \\lambda_3 \\mathcal{L}_{dr} + \\lambda_4 \\mathcal{L}_{fea}\".","section":"Sec. 4.3, Table 1"},{"comment":"The headline comparison in Table 1 uses different initialization strategies for the baselines (SfM points from three views for 2DGS/GOF/PGSR) versus the proposed method (CLMVSNet dense points). The controlled comparison in Table 3 (2DGS+CLMVSNet: 1.27, PGSR+CLMVSNet: 1.38, Ours: 1.13) is more informative for isolating the contribution of the proposed losses; please reference it directly in the main text near Table 1 to avoid overstating the margin against 2DGS.","section":"Sec. 3.2"},{"comment":"After Eq. (3), please clarify how the feature maps F_i are resized or interpolated when querying at fractional pixel coordinates \\pi(p_{ij}), and whether the same FPN resolution (8 channels) is used for both the feature splatting loss and the DGPR cross-view consistency loss.","section":"Sec. 3.4"},{"comment":"The description of the selective update does not specify the patch size used for NCC, the number of update iterations over the 7000 training steps, or whether the update is applied to all Gaussians or only those with sufficient gradient. Please add these implementation details for reproducibility.","section":"Sec. 4.4"},{"comment":"Table 4's header contains a typo: \"Compleness\" should be \"Completion\".","section":"Sec. 4.3"},{"comment":"The claim that the method is \"2x faster than the NeRF-based fine-tuning approach\" is based on comparison with SparseNeuS (Table 7); consider stating this explicitly in the abstract and main text, since NeuSurf's training time is much larger but its CD is better (0.99).","section":"Sec. 4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a well-executed incremental contribution with code release and a standard benchmark. The main risk is that the DGPR improvement is an artifact of the fixed MVS features rather than a genuine geometric cue, and that the single-run ablations overstate the per-module gains. The authors should be asked to provide variance estimates and feature-consistency validation. I do not see a fundamental circularity problem: the evaluation metric is external ground-truth Chamfer distance, and the final result improves on the MVS initialization alone. The paper's fit to the journal is acceptable if the revisions are substantive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nSparse2DGS is a decent empirical paper, not a breakthrough. The headline is that initializing 2DGS with an unsupervised MVS point cloud plus three geometry-focused tricks gets you from 2.81 CD down to 1.13 on the DTU 3-view benchmark, beating the Gaussian baselines (GOF 2.82, PGSR 2.08) and matching or beating the older NeRF-based methods like SparseNeuS (1.27) at lower training cost. The 2x speedup over SparseNeuS is real. The novelty is modest but present: nobody else has combined MVS initialization with fixed-color/feature splatting and a reparameterized disk-sampling regularizer in quite this way.\n\nWhat the paper does well: the ablation set is thorough. Each of the four modules is toggled, and the standard DTU protocol is followed. The reparameterization trick (sampling points on the Gaussian disk, then back-propagating through the sampling to orientation and scale) is a clean idea and worth stealing. The code is available. The paper's central claim—that the full pipeline beats both 2DGS and CLMVSNet alone—holds up in the reported numbers.\n\nSoft spots, in proportion. First, all experiments are single runs, no seeds, no error bars. On a benchmark where 0.05 CD can be noise, several ablation margins (1.201 vs 1.142 for DGPR) are uncomfortably thin. Second, the abstract overstates: 'outperforms existing methods by notable margins' is not accurate, because NeuSurf reports 0.99 CD in the same table. The comparison should be framed as best among Gaussian-based methods and competitive with NeRF-based ones. Third, the stress-test worry about DGPR is real and unaddressed. The loss assumes CLMVSNet FPN features are locally view-consistent at arbitrary points sampled off the Gaussian disk; the paper never checks this, doesn't mask occluded samples, and its own limitation section concedes occlusion errors. The gain from DGPR is small enough that it could be fitting spurious feature similarity rather than true geometry. I'd want the authors to add an analysis of when DGPR helps/hurts, or at least error bars and an occlusion mask.\n\nWho is this for? Anyone building sparse-view Gaussian splatting systems will find the MVS initialization and the sampling trick useful. It deserves a serious referee; with the single-run issue and the abstract fixed, it could be a solid conference paper. My recommendation: send it out, but ask for seeds, a corrected comparison, and a deeper look at DGPR.","headline":"A solid, well-ablated empirical paper on sparse-view Gaussian splatting that deserves peer review, though the single-run numbers and an overstateful abstract need fixing.","tokens_in":13997,"tokens_out":2795,"would_cite":true,"duration_ms":27056,"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":"Sparse2DGS shows that seeding 2D Gaussian splatting with an MVS point cloud and prioritizing geometry yields accurate surfaces from three views.","keywords":["sparse-view 3D reconstruction","Gaussian splatting","surface reconstruction","multi-view stereo","2D Gaussian primitives","cross-view consistency","geometric regularization","DTU dataset"],"falsifier":"On a sparse-view set with a large occluded region, measure per-region Chamfer error against ground truth for Sparse2DGS and for the plain 2DGS-plus-CLMVSNet baseline; if the occluded regions do not improve, or if deliberately corrupting CLMVSNet depth in those regions leaves the overall error essentially unchanged, then the geometric-prioritized schemes are not doing the corrective work the paper attributes to them.","tokens_in":12922,"feed_emoji":"📐","tokens_out":6847,"duration_ms":65263,"temperature":0.7,"pith_summary":"Sparse2DGS sets out to show that accurate and complete 3D surface reconstruction is possible from as few as three posed images, by seeding 2D Gaussian Splatting with a learning-based multi-view stereo (MVS) point cloud and then steering optimization toward geometry rather than appearance. The paper reports that a plain MVS-plus-2DGS combination fails to beat the MVS points alone (Chamfer distance 1.27 vs 1.26), but that adding three geometry-prioritized schemes—fixed-color MVS feature supervision, reparameterization-based disk regularization, and selective Gaussian updates from rendered depth and normals—lowers the DTU 3-view Chamfer distance to 1.13, below all compared Gaussian splatting methods. If this holds, sparse-view reconstruction becomes practical for scenes where only a handful of images exist, and the approach is fast enough (about 10 minutes) to serve as a practical test-time pipeline.","feed_headline":"MVS-seeded splatting rebuilds surfaces from 3 views","feed_subtitle":"Chamfer error drops to 1.13 on DTU versus 2.81 for 2DGS, with training in about 10 minutes.","key_machinery":"The mechanism that carries the argument is a reparameterized 2D Gaussian disk: points are drawn from a standard Gaussian and transformed by $Z = p + RSZ'$, where $R$ is the disk orientation, $S$ its scale, and $p$ its position, so all three geometric properties receive gradients from a single point-based cross-view feature consistency loss. This direct Gaussian primitive regularization is backed by geometrically enhanced supervision, in which CLMVSNet FPN features and RGB colors are splatted into the renderer and held fixed, with rendered features supervised by cosine similarity, and by selective Gaussian update, in which patch-wise normalized cross-correlation over rendered depth and normal maps decides when to reproject a point and replace a primitive position. Together these convert a Gaussian primitive into a set of query points that can be checked for multi-view agreement, while keeping appearance from absorbing the sparse-view ambiguity.","core_discovery":"The central claim is that the main obstacle to sparse-view Gaussian surface reconstruction is not missing primitives but missing geometric constraint, and that MVS-derived points plus features can supply that constraint if geometry is prioritized. Sparse2DGS initializes every 2D disk primitive from back-projected CLMVSNet depth, freezes per-primitive color and MVS FPN feature values so appearance cannot overfit the few views, represents each disk's position, orientation, and scale through sampled points via $Z = p + RSZ'$ supervised by cross-view feature consistency, and replaces adaptive densification with a selective update that repositions a primitive when patch-wise NCC favors the rendered depth/normal cues. On the DTU three-view benchmark the paper reports a mean Chamfer distance of 1.13, compared with 2.81 for 2DGS, 2.82 for GOF, and 2.08 for PGSR, with training time about 10 minutes—2x faster than fine-tuning SparseNeuS and roughly 60x faster than NeuSurf.","pith_inferences":["The same recipe—a dense geometric prior that supplies both points and features, frozen appearance, and point-based geometric regularization—could plausibly work with other priors such as monocular depth or foundation-model features, though the paper only tests CLMVSNet.","Because the cross-view consistency loss uses features extracted once, the method's ceiling is largely set by the discriminative power and view consistency of those features; improving or fine-tuning the feature extractor should directly lower reconstruction error, a testable prediction.","The selective update step is a local geometric refinement; running it more often or combining it with visibility checks might extend the method to scenes with larger occlusions, where the paper's stated limitation predicts degraded performance.","If the reported DTU margin is robust, the practical implication is that sparse-view surface reconstruction no longer requires per-scene NeRF fine-tuning or hours of optimization, which should make three-image 3D capture pipelines substantially cheaper."],"forward_implications":["On the three-view DTU benchmark, Sparse2DGS reports mean Chamfer distance 1.13, below 2DGS (2.81), GOF (2.82), and PGSR (2.08).","The method reconstructs more complete surfaces in qualitative comparisons on DTU and BlendedMVS, preserving scene structure where prior Gaussian methods produce noise and missing regions.","Training finishes in about 10 minutes, roughly 2x faster than SparseNeuS fine-tuning and about 60x faster than NeuSurf, making the approach practical as a test-time pipeline.","Ablations show that fixing color and MVS feature values helps geometry more than lowering spherical-harmonic degrees, supporting the paper's diagnosis of appearance overfitting under sparse views.","Replacing adaptive density control with selective Gaussian update avoids the degradation of densified primitives under sparse supervision in the paper's ablation (1.152 vs 1.234 Chamfer distance)."],"supporting_citations":[{"why":"Provides the dense MVS point cloud used for Gaussian initialization and the FPN feature maps used for geometrically enhanced supervision.","marker":"[31]"},{"why":"Supplies the 2D disk primitive representation, differentiable renderer, and base image, depth-normal, and depth-distortion losses that Sparse2DGS extends.","marker":"[11]"},{"why":"The original Gaussian splatting framework whose adaptive density control is compared against and replaced by the selective Gaussian update.","marker":"[14]"},{"why":"The NeRF-based sparse-view baseline used for the speed comparison (25 minute fine-tuning vs about 10 minutes).","marker":"[20]"},{"why":"A planar Gaussian splatting baseline that Sparse2DGS reports outperforming on DTU (2.08 vs 1.13 Chamfer distance).","marker":"[4]"},{"why":"A Gaussian opacity-field baseline used in the DTU three-view comparison (2.82 Chamfer distance).","marker":"[43]"},{"why":"Introduces the cross-view feature consistency loss that the reparameterized disk sampling feeds into.","marker":"[45]"},{"why":"The DTU dataset and official Chamfer Distance evaluation protocol on which all quantitative claims are measured.","marker":"[1]"}],"fun_headline_variants":["Sparse2DGS: Geometry-prioritized splatting from sparse views","Sparse2DGS: 3-view DTU error drops to 1.13 in 10 minutes","Geometry-prioritized splatting beats MVS on sparse views","Sparse2DGS: 10-minute splatting for sparse-view surfaces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the MVS feature maps, computed once from the three input views, are locally view-consistent at sub-pixel positions for arbitrary points sampled on the Gaussian disks—points that generally do not lie exactly on surfaces—and that the MVS depths themselves remain reliable in occluded and textureless regions; the paper's own limitation section concedes that occlusion errors from MVS can propagate and hurt performance.","fun_headline_variants_meta":{"raw":{"variants":["Sparse2DGS: Geometry-prioritized splatting from sparse views","Sparse2DGS: 3-view DTU error drops to 1.13 in 10 minutes","Geometry-prioritized splatting beats MVS on sparse views","Sparse2DGS: 10-minute splatting for sparse-view surfaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000499,"raw_usage":{"total_tokens":2428,"prompt_tokens":915,"completion_tokens":1513,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":1421}},"tokens_in":531,"tokens_out":1513,"duration_ms":11587,"temperature":1.0,"reasoning_tokens":1421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:31:12.587294+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a sparse-view set with a large occluded region, measure per-region Chamfer error against ground truth for Sparse2DGS and for the plain 2DGS-plus-CLMVSNet baseline; if the occluded regions do not improve, or if deliberately corrupting CLMVSNet depth in those regions leaves the overall error essentially unchanged, then the geometric-prioritized schemes are not doing the corrective work the paper attributes to them.","supporting_citations":[{"cited_title":"Cl-mvsnet: Unsupervised multi-view stereo with dual-level contrastive learning","cited_arxiv_id":null,"evidence_quote":"Provides the dense MVS point cloud used for Gaussian initialization and the FPN feature maps used for geometrically enhanced supervision."},{"cited_title":"2d gaussian splatting for geometrically ac- curate radiance fields","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D disk primitive representation, differentiable renderer, and base image, depth-normal, and depth-distortion losses that Sparse2DGS extends."},{"cited_title":"Sparseneus: Fast generalizable neural sur- face reconstruction from sparse views","cited_arxiv_id":null,"evidence_quote":"The NeRF-based sparse-view baseline used for the speed comparison (25 minute fine-tuning vs about 10 minutes)."},{"cited_title":"Learning signed distance field for multi-view surface reconstruction","cited_arxiv_id":null,"evidence_quote":"Introduces the cross-view feature consistency loss that the reparameterized disk sampling feeds into."},{"cited_title":"Large-scale data for multiple-view stereopsis","cited_arxiv_id":null,"evidence_quote":"The DTU dataset and official Chamfer Distance evaluation protocol on which all quantitative claims are measured."}],"review_version":1}