{"id":"6102d737-197e-4bcd-811b-cead430e4fd0","arxiv_id":"2505.19854","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using DUSt3R and COLMAP MVS point clouds to initialize 2D Gaussians improves three-view surface reconstruction accuracy on the DTU dataset.","lead":"Sparse2DGS builds a 3D model from just three photos by feeding 2D Gaussian Splatting with dense point clouds from DUSt3R and COLMAP MVS, instead of the sparse points that structure from motion usually gives. It reports lower reconstruction error than several baselines on the DTU benchmark and runs in about 100 seconds per scene, so it is worth reading for anyone building fast sparse-view 3D capture.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation in Table 2 conflates post-processing with MVS integration: the missing 'DUSt3R (Ours) only' row leaves open that the 0.25 mean-CD gain comes from downsampling and outlier removal alone, not from the proposed integration.","rationale":"The reader's weakest-assumption is that ICP alignment between the DUSt3R and COLMAP MVS point clouds may fail (Sec. 3.2). That is a legitimate concern about reliability, and it is unresolved because no alignment error or failure cases are reported. However, the current experiments already show that the integrated initialization yields good results on the 12 reported scenes, so ICP appears to work at least on those scenes. The more fundamental gap is that the ablation does not isolate the contribution of MVS integration at all: the comparison is between raw DUSt3R and post-processed DUSt3R plus MVS. Since the post-processing is also part of the proposed pipeline, the improvement from 1.76 to 1.51 may be attributable to downsampling and outlier removal rather than to adding MVS points. If that were the case, the paper's stated contribution, integrating DUSt3R with COLMAP MVS, would be unsupported, even though the broader idea of using a dense point cloud to initialize 2DGS might still be valid. I therefore disagree partially with the reader's focus: the missing ablation control is more load-bearing than the unvalidated ICP, because it determines whether the integration step has any measured benefit. The proposed concrete test is a single additional ablation row that settles this directly. I keep the verdict at CONDITIONAL (i.e., unchanged from the reader) because the issue is correctable and does not invalidate the method outright, but it should be a required revision before acceptance.","tokens_in":8940,"tokens_out":13980,"duration_ms":143219,"concrete_test":"Add an ablation row to Table 2: initialize 2DGS with the post-processed DUSt3R cloud only (i.e., DUSt3R (Ours) without MVS) on the same 12 selected DTU scenes, and report the mean and per-scan Chamfer Distance alongside the existing rows. If the mean is within 0.05 of the proposed method's 1.51, the MVS integration contributes little; if it is above 1.61, the integration is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adding COLMAP MVS points to a post-processed DUSt3R cloud improves 2DGS initialization. Table 2 does not test this claim cleanly. The row labeled 'DUSt3R' (1.76) initializes 2DGS with the raw DUSt3R output, while the row 'MVS + DUSt3R (Ours)' (1.51) uses the downsampled, outlier-filtered DUSt3R cloud merged with MVS via ICP. The table has no row for the post-processed DUSt3R cloud alone, even though the caption states that 'DUSt3R (Ours)' is the post-processed version. Consequently, the observed 0.25 mean-CD improvement could be caused entirely by the post-processing described in Sec. 3.2 (voxel downsampling and statistical outlier removal), rather than by the MVS integration that the paper highlights. Without this control, the experiments do not demonstrate that the integration step, or the ICP alignment it requires, contributes to the result. Reporting the missing ablation row, ideally with per-scan values and variance, is necessary to support the paper's central mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Sparse2DGS, a method for surface reconstruction from three input images. The method uses DUSt3R to generate a dense point cloud, post-processes it with voxel downsampling and statistical outlier removal, aligns it to a COLMAP MVS point cloud via ICP, and uses the merged cloud to initialize 2D Gaussian Splatting without densification. Surface meshes are obtained by rendering depth maps and fusing them with TSDF integration. On 12 selected DTU scenes, the method reports a mean Chamfer Distance of 1.51, matching UFORecon and improving over 2DGS (2.54) and raw DUSt3R initialization (2.09/1.76 depending on protocol), with about 100 seconds per scene.","tokens_in":9163,"tokens_out":4502,"duration_ms":49374,"significance":"If the reported results are reproducible, Sparse2DGS would be a practically attractive sparse-view surface reconstruction method: it avoids the multi-day pretraining of methods such as UFORecon, requires only three images, and reaches comparable mean Chamfer Distance on the chosen DTU subset. The paper's internal ablation in Table 2 shows a clear improvement of 2DGS with different initialization sources, and the qualitative results suggest meaningful surface completeness gains on difficult scans such as scan 24. However, the central mechanism — that integrating COLMAP MVS points with post-processed DUSt3R points is what drives the improvement — is not isolated by the reported ablation, and the benchmark protocol lacks the specificity needed to reproduce the exact numbers.","major_comments":[{"comment":"","section":"§4.5, Table 2"},{"comment":"","section":"§3.2"},{"comment":"","section":"§4.1, Table 1"}],"minor_comments":[{"comment":"","section":"§1"},{"comment":"","section":"References"},{"comment":"","section":"§5"},{"comment":"","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The experimental design is the main weakness. The missing ablation row in Table 2 is a straightforward fix but is essential: without it, the paper's central contribution is not isolated. The benchmark protocol also needs tightening. I saw no evidence of problematic citation practice. The paper is compact and likely intended for a workshop-style venue; the requested changes are within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper reports a simple, fast pipeline for three-view surface reconstruction by initializing 2DGS with a merged DUSt3R + COLMAP MVS point cloud. The idea is straightforward and likely useful, but the main ablation does not actually isolate the integration step from the post-processing, and the comparison is missing some key baselines.\n\nWhat's new: the specific combination of DUSt3R and COLMAP MVS point clouds as Gaussian initialization, with a 2DGS backbone, and a quantitative result on 12 DTU scans. The speed (~100s) is attractive. The paper is honest enough to show an ablation where MVS alone and DUSt3R alone each beat COLMAP SfM, and the full method gets the best mean CD. That is reproducible and useful.\n\nThe stress-test concern is correct. Table 2's caption says 'DUSt3R (Ours)' is the post-processed version, but the table only shows the post-processed DUSt3R in the combined row, never alone. So you cannot tell whether the 0.25 mean-CD gain from 1.76 to 1.51 comes from adding COLMAP MVS points or just from voxel downsampling and outlier removal. This is a load-bearing missing control for the paper's central mechanism.\n\nOther issues: reflective scans (69, 97, 110) are excluded, which narrows the claim; view selection is 'difficult' without a reproducible protocol (citing SPARF); UFORecon is evaluated without its random-set training strategy, possibly handicapping it; no variance or code; and pixelSplat/MVSplat are not cited, so the novelty is not sharply positioned. These are correctable.\n\nBalanced view: the core empirical result—that dense point cloud initialization helps 2DGS from three views—is plausible and the paper shows it. The weakness is in the attribution, not in the motivating problem. The paper is short (workshop length), and the missing row is a small experimental fix.\n\nI would send it to peer review, not desk reject. The ablation gap should be flagged as major, but the pipeline is simple enough that one missing row and a better baseline comparison could make the claim stand. I'd want the missing 'DUSt3R (Ours) alone' row, per-scan values, and ideally code. For a reading group, it's a maybe—useful as an example of a practical sparse-view system, not as a methodological breakthrough.","headline":"A plausible sparse-view 2DGS initialization trick that is useful but under-supported by the current ablation, which omits the isolation needed to attribute the gain to the MVS integration.","tokens_in":9736,"tokens_out":2352,"would_cite":false,"duration_ms":22223,"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 reconstructs accurate 3D surfaces from just three images by initializing 2D Gaussian splatting with a dense point cloud.","keywords":["sparse-view 3D reconstruction","2D Gaussian splatting","DUSt3R","COLMAP MVS","point cloud initialization","surface reconstruction","Gaussian splatting","three-view reconstruction"],"falsifier":"Run the pipeline on a scene pair where the DUSt3R and COLMAP point clouds are known to be misaligned, for example a highly reflective object, and measure the Chamfer Distance; if the error jumps to or above the DUSt3R-only baseline, the alignment step is the failure point. A more direct test is to replace the ICP alignment with ground-truth alignment of the two clouds and check whether the mean Chamfer Distance improves; if it does not, misalignment is not the limiting factor.","tokens_in":8724,"feed_emoji":"🧊","tokens_out":7767,"duration_ms":66080,"temperature":0.7,"pith_summary":"The paper claims that the main obstacle to sparse-view Gaussian splatting is the sparse, unreliable point cloud that seeds the Gaussians. It proposes to replace that seed with an integrated dense point cloud built from DUSt3R's per-pixel stereo predictions and COLMAP's multi-view stereo output. With this initialization, 2D Gaussian splatting can reconstruct object surfaces from just three images, reaching a mean Chamfer Distance of 1.51 on twelve DTU scenes, matching a state-of-the-art sparse-view method that requires days of pre-training, in about 100 seconds per scene. The central load-bearing claim is that the fused dense point cloud supplies enough correct Gaussian centers to keep optimization from collapsing when views are few.","feed_headline":"Sparse2DGS reconstructs accurate 3D surfaces from just three images","feed_subtitle":"Fusing DUSt3R and COLMAP point clouds to seed Gaussians matches a pretrained method's accuracy in 100 seconds.","key_machinery":"The central object is the 2D Gaussian primitive from 2D Gaussian splatting: an ellipse lying in the local tangent plane of the surface, defined by a center, two tangent vectors, two scale factors, opacity, and color. The paper's carrying mechanism is the integrated dense point cloud that initializes these centers. DUSt3R produces a per-pixel point map for each input view; the paper downsamples it with voxels of size 0.005, removes statistical outliers, and aligns it to the COLMAP MVS point cloud via iterative closest point (ICP) registration. Each fused point becomes the initial position of a 2D Gaussian, so no densification is performed during optimization. The 2DGS losses—color reconstruction, depth distortion, and normal consistency—then optimize the Gaussians, and the rendered depth maps are merged by truncated signed distance function (TSDF) integration into a mesh.","core_discovery":"The discovery is that dense, accurate initialization is what makes Gaussian splatting viable in the sparse-view regime, and that the fusion of DUSt3R and COLMAP MVS provides exactly that dense initialization. The paper shows that using only the sparse COLMAP SfM point cloud, as in standard 2DGS, yields a mean Chamfer Distance of 2.54; using COLMAP MVS alone gives 1.90; using the post-processed DUSt3R cloud alone gives 1.76; and fusing both gives 1.51. The method does not densify Gaussians during optimization; it optimizes the parameters of the Gaussians placed at the dense initial points, and renders depth maps that are integrated into a mesh via TSDF fusion.","pith_inferences":["The same fused-initialization idea could be dropped into other Gaussian-splatting reconstruction methods, for example 3DGS-based surface extractors, to test whether the dense seed helps beyond the 2DGS formulation.","The reliance on ICP suggests a robustness bottleneck: a learned or globally consistent alignment, such as the one DUSt3R can produce for stereo pairs, might remove the need for a separate alignment step and make the method work on scenes where ICP is prone to local optima.","The method's success with three views hints that the key difficulty in sparse-view reconstruction is initialization density rather than the optimization objective itself; if so, even simpler reconstruction losses could benefit from the same dense seed.","A natural stress test would be running the method with only two input images, where the DUSt3R stereo prior is stronger and COLMAP MVS is weaker, to see whether the fusion still holds up."],"forward_implications":["Three-image reconstruction is now practical without pre-training, since the method runs in about 100 seconds per scene on a single GPU.","The ablation quantifies the contribution of each point cloud source: COLMAP MVS alone lowers the mean Chamfer Distance to 1.90, post-processed DUSt3R alone to 1.76, and their fusion to 1.51.","Because no densification is used, the number of Gaussians is fixed by the fused point cloud, making the optimization faster and more deterministic than standard Gaussian splatting.","The method matches the accuracy of a pre-trained sparse-view neural surface method while requiring only three images, ground-truth camera parameters, and no offline training for each scene."],"supporting_citations":[{"why":"Supplies the dense per-pixel point maps that form the primary initialization for Gaussians.","marker":"[12]"},{"why":"Provides the second dense point cloud via COLMAP MVS, fused with the DUSt3R cloud, and serves as an MVS baseline.","marker":"[2]"},{"why":"Defines the 2D Gaussian representation, rendering equations, and the depth distortion and normal consistency losses that the method optimizes.","marker":"[10]"},{"why":"The iterative closest point algorithm used to align the DUSt3R point cloud to the COLMAP MVS point cloud before fusion.","marker":"[21]"},{"why":"Provides the standard sparse SfM initialization that the ablation compares against and that the method replaces.","marker":"[15]"},{"why":"Defines the base 3D Gaussian splatting framework and the color reconstruction loss used in the total loss.","marker":"[9]"},{"why":"DTU dataset provides the twelve scenes, three-view selections, ground-truth camera parameters, and ground-truth point clouds used for evaluation.","marker":"[13]"}],"fun_headline_variants":["3D reconstruction from three photos via dense point seeding","Just three images yield accurate 3D surfaces with dense seeding","Sparse-view 3D surfaces: three images, dense point clouds","Three-image 3D reconstruction boosted by dense point clouds","Accurate 3D from three images via dense Gaussian seeding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that the ICP alignment between the DUSt3R point cloud and the COLMAP MVS point cloud converges to the correct global alignment, so that the fused Gaussian centers actually lie on the object surface.","fun_headline_variants_meta":{"raw":{"variants":["3D reconstruction from three photos via dense point seeding","Just three images yield accurate 3D surfaces with dense seeding","Sparse-view 3D surfaces: three images, dense point clouds","Three-image 3D reconstruction boosted by dense point clouds","Accurate 3D from three images via dense Gaussian seeding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000499,"raw_usage":{"total_tokens":2446,"prompt_tokens":952,"completion_tokens":1494,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":1409}},"tokens_in":568,"tokens_out":1494,"duration_ms":10387,"temperature":1.0,"reasoning_tokens":1409,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:05:25.752017+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a scene pair where the DUSt3R and COLMAP point clouds are known to be misaligned, for example a highly reflective object, and measure the Chamfer Distance; if the error jumps to or above the DUSt3R-only baseline, the alignment step is the failure point. A more direct test is to replace the ICP alignment with ground-truth alignment of the two clouds and check whether the mean Chamfer Distance improves; if it does not, misalignment is not the limiting factor.","supporting_citations":[{"cited_title":"Szeliski, Computer Vision: Algorithms and Applications , Springer-V erlag New Y ork Inc., 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the dense per-pixel point maps that form the primary initialization for Gaussians."},{"cited_title":"Sparse2DGS: Sparse-View Surface Reconstruction using 2D Gaussian Splatting with Dense Point Cloud","cited_arxiv_id":"2505.19854","evidence_quote":"Provides the second dense point cloud via COLMAP MVS, fused with the DUSt3R cloud, and serves as an MVS baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the 2D Gaussian representation, rendering equations, and the depth distortion and normal consistency losses that the method optimizes."},{"cited_title":"SuGaR: Surface-aligned gau ssian splatting for efficient 3D mesh reconstruction and high-qua lity mesh rendering,","cited_arxiv_id":null,"evidence_quote":"The iterative closest point algorithm used to align the DUSt3R point cloud to the COLMAP MVS point cloud before fusion."},{"cited_title":"NeRF: Respresenting scenes as neural radiance fields for view synthesis,","cited_arxiv_id":null,"evidence_quote":"Provides the standard sparse SfM initialization that the ablation compares against and that the method replaces."},{"cited_title":"In Sparse2DGS, we utilize DUSt3R [12] and COLMAP MVS [2] to obtain the dense 3D point clouds for initializing 2D Gauss ians in 2DGS","cited_arxiv_id":null,"evidence_quote":"Defines the base 3D Gaussian splatting framework and the color reconstruction loss used in the total loss."},{"cited_title":"Pix- elwise view selection for unstructured multi-view stereo,","cited_arxiv_id":null,"evidence_quote":"DTU dataset provides the twelve scenes, three-view selections, ground-truth camera parameters, and ground-truth point clouds used for evaluation."}],"review_version":1}