{"id":"64a34fb5-7c42-4de1-9e48-e0b7e1c6d3fe","arxiv_id":"2501.01003","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EasySplat combines DUSt3R pointmap initialization with a KNN-based Gaussian splitting rule to improve 3D Gaussian Splatting for dense-view novel view synthesis.","lead":"EasySplat replaces the usual SfM-based initialization of 3D Gaussian Splatting with a grouping strategy over DUSt3R pointmap priors and adds a KNN-based densification rule. The authors report higher average PSNR and LPIPS than previous methods on Tanks&Temples, though the claim of best across all metrics is not fully supported.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The grouping heuristic underpinning the initialization contribution is validated on only one scene; matching all non-reference frames to a single reference per group can create low-overlap pairs that DUSt3R cannot constrain, and Table 1 shows the pose story is actually mixed.","rationale":"Reader's weakest assumption is the grouping heuristic, and I agree that it is load-bearing. The refinement here is that the unvalidated step is not just where cuts are placed but the one-reference-per-group design in Eq. (3), which is independent of cut placement and worsens with group length. The paper's own Table 1 shows the pose benefit is not clean: ATE is worse than oneref, so the narrative 'more accurate camera poses' is not supported by the reported numbers; the PSNR gain may come from denser initial point cloud or from 3DGS training, but the paper does not disentangle this. I considered the undefined matrix inequality in the KNN densification (Σ3di > ¯Σ3d in Eq. 7) as an alternative, and it is a real reproducibility gap, but it does not threaten the conceptual claim as directly because some scalar reduction (trace or determinant) is likely used. I also considered the SSIM overclaim, which is real but narrow: it weakens 'best across all metrics' without overturning the PSNR/LPIPS claim. The grouping/reference-overlap concern therefore remains the most load-bearing: if it fails, the initialization contribution fails and the SOTA numbers are dataset-specific. The concrete test above would settle it. Since the concern is substantial but testable and the reported methodology is otherwise coherent, keeping the reader's CONDITIONAL verdict is appropriate.","tokens_in":9259,"tokens_out":8569,"duration_ms":85247,"concrete_test":"On the eight Tanks&Temples sequences, compute for each cosine-similarity group the reference-to-farthest-frame overlap (e.g., DUSt3R confidence or angular baseline). Then run EasySplat with two ablations: (a) the paper's k=2 one-reference-per-group scheme, and (b) a bounded-overlap baseline that introduces a new reference whenever the reference-to-frame overlap falls below a fixed threshold, using the same DUSt3R and 3DGS code. Compare PSNR, LPIPS, ATE, and RPE on the same test split. If (b) matches or beats (a), the single-reference-per-group constraint is the active weakness; if (b) is worse, the heuristic is adequate and the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"EasySplat's first contribution reduces to the claim that the Eq. (2) cosine-difference cuts produce groups in which every non-reference view has enough overlap with the group's reference view for DUSt3R pointmaps and Eq. (4) global alignment to be reliable. This is load-bearing because the rest of the pipeline is initialized from that point cloud and those poses, and Table 1 is the only direct evidence. The evidence is thin and partly ambiguous: on Church, 'ours' improves PSNR over oneref (30.22 vs 29.80) but has worse ATE (0.005 vs 0.003), so the gain is not cleanly attributable to better camera poses. More importantly, Eq. (3) fixes one reference per group, and k=2 is fixed for all sequences, so group length grows with sequence length without any bound on reference-to-frame overlap. DUSt3R confidence-weighted alignment in Eq. (4) down-weights low-overlap pairs, but then the affected frames are weakly constrained in both pose and geometry; 3DGS training cannot recover from a bad global initialization. No experiment varies k, group geometry, or sequence length, and no error bars are reported. If this assumption fails on repetitive-texture or large-motion sequences, the reported SOTA numbers would not generalize.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EasySplat, a 3D Gaussian Splatting framework that replaces SfM initialization with a DUSt3R-based pointmap prior using a view-similarity grouping strategy, and replaces the standard densification rule with a KNN-based adaptive splitting criterion. The authors report state-of-the-art novel view synthesis results on Tanks&Temples and CO3DV2, along with ablations for the two proposed components.","tokens_in":9606,"tokens_out":3825,"duration_ms":35152,"significance":"If the reported results are reliable, the paper would make a useful contribution by showing that a pointmap-based prior can replace SfM for dense-view 3DGS initialization and that a local-shape-aware densification rule improves rendering quality. The idea of view-adaptive grouping to make pairwise pointmap estimation tractable on dense sequences is reasonable and potentially valuable. However, the current evidence is not fully consistent: the SSIM result in the main comparison does not support the 'best across all metrics' claim, the pose-accuracy claims are contradicted by part of the reported CO3DV2 numbers, and the grouping heuristic is validated on only one scene. The paper does not provide code, repeated runs, or error bars, which limits how strongly the conclusions can be drawn.","major_comments":[{"comment":"The text states that 'our method achieves the best performance across all metrics,' but Table 2 does not support this. ScaffoldGS has a higher mean SSIM (0.956 vs. 0.953) and wins on four scenes (Church, Barn, Francis, Ignatius) and ties on two (Museum, Ballroom). The claimed SOTA is therefore only true for mean PSNR and LPIPS. Please revise the claim and discuss the SSIM trade-off explicitly, or provide an explanation for why the average SSIM is lower despite higher PSNR and LPIPS.","section":"§IV.B, Table 2"},{"comment":"The view-adaptive grouping strategy is load-bearing because all downstream results depend on the resulting point cloud and poses, yet its evaluation is limited to a single scene (Church). On that scene, the proposed grouping has worse ATE than the oneref baseline (0.005 vs. 0.003) while improving PSNR, so the gain cannot be cleanly attributed to more accurate camera poses. No experiment varies k, group size, sequence length, or scene content, and no error bars are reported. The fixed choice k=2 for all sequences leaves open the risk that groups become too long for the reference-to-frame overlap to be sufficient. Please provide per-scene grouping statistics and a sensitivity analysis of k, and verify the pose-quality claim on more scenes.","section":"§III.B, Table 1 and Eqs. (2)–(4)"},{"comment":"The proposed densification condition is not well defined. Eq. (6) compares the covariance matrix Σ3di with the scalar threshold τS, and the subsequent text states 'If Σ3di > ¯Σ3d' comparing two matrices. No partial order on covariance matrices is defined. If the intended comparison is by determinant, trace, or largest eigenvalue, this must be stated explicitly; as written, the method is not reproducible. This is a central part of the second contribution, not just a notation issue.","section":"§III.C, Eqs. (6)–(7)"},{"comment":"The CO3DV2 evaluation uses only four scenes, chosen randomly but without listing the scene names, and no repeated runs or confidence intervals are reported. More importantly, the statement that EasySplat 'significantly outperforms CF-3DGS in both novel view synthesis and camera pose estimation' is contradicted by the table: EasySplat has worse RPE t than CF-3DGS on all four scenes (e.g., 0.592 vs. 0.257, 0.427 vs. 0.112) and mixed RPE r results. The claim should be restricted to ATE and NVS metrics, and the pose-error discrepancy should be discussed.","section":"§IV.B, Table 3"}],"minor_comments":[{"comment":"Figure 2 labels the pointmap model as 'MASt3R', while the text describes using 'DUSt3R's pretrained model' and reference [34] is the MASt3R paper. Please align the notation so that the reader knows which pretrained model was actually used.","section":"Figure 2 and §III.B"},{"comment":"Several references have corrupted or incomplete author lists, most notably [1] ('Bernhard Müller, Georgios Kerbl, ...' for the 3DGS paper) and [3] (a garbled list starting 'Yang, Sifei Fu, Amey Liu, ...'). These should be corrected to their proper bibliographic entries.","section":"References"},{"comment":"The 'swin' scheme is not defined anywhere in the text. If it refers to a sliding-window pairing strategy, please define it explicitly so that the comparison in Table 1 is interpretable.","section":"Table 1"},{"comment":"The cosine similarity in Eq. (1) treats images as flattened vectors, which implicitly requires identical dimensions and does not account for spatial structure. Please state this assumption explicitly.","section":"Eq. (1)"},{"comment":"A stray Chinese-language heading, '六、策略二消融可视化对比', appears immediately before Figure 3 and should be removed or translated.","section":"Figure 3 vicinity"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core idea, but the current numerical evidence is more mixed than the text claims. I would ask the authors to correct the SOTA claim regarding SSIM, provide a more thorough validation of the grouping strategy, define the covariance comparison, and temper the CO3DV2 pose claims. These are fixable with additional experiments and careful rewriting, so I do not recommend rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chen —\n\nEasySplat replaces COLMAP with DUSt3R pointmaps for 3DGS initialization and adds a KNN-based densification rule that splits large Gaussians based on average neighbor shape. The KNN rule is a simple, testable idea that could be adopted by others; the ablations in Table 4 show it helps. The view-adaptive grouping—cut the sequence at the largest cosine-similarity differences and match all frames to one reference per group—is a reasonable way to keep DUSt3R tractable in dense-view settings, and Table 1 shows it beats a single-reference baseline in PSNR.\n\nThe paper overclaims. Table 2 has ScaffoldGS with higher mean SSIM (0.956 vs 0.953) and wins on four of eight scenes, so \"best across all metrics\" is false. The mean PSNR and LPIPS gains are real but modest.\n\nThe weaker spot is the grouping itself. k=2 is fixed for every sequence, and each group uses a single reference. For long videos or large viewpoint jumps, some non-reference frames will have little overlap with the reference, and DUSt3R's confidence weighting will down-weight them—leaving pose and geometry weakly constrained. That's load-bearing because everything downstream depends on this initialization. The only direct evidence is Church (Table 1), where ATE is actually worse than oneref (0.005 vs 0.003) even though PSNR improves. No sensitivity analysis on k, no variation of group geometry, no error bars. CO3DV2 is four scenes with no names and no repeated runs. No code is released.\n\nNone of this is fatal. The ideas are plausible and the paper is clearly written. It needs a de-overclaimed abstract, error bars or repeated runs, a sensitivity study on k and n, and code. As is, I'd send it to review, but I'd expect the revision to be substantial.\n\nFor you: relevant if you work on 3DGS initialization or densification. Not urgent.","headline":"Plausible DUSt3R-based initialization and a simple KNN densification rule, but the SOTA claim is overstated and the grouping heuristic is under-validated.","tokens_in":10118,"tokens_out":3327,"would_cite":false,"duration_ms":29836,"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":"EasySplat claims that two targeted changes to 3D Gaussian Splatting — a view-similarity grouping rule for pointmap initialization and a nearest-neighbor-based splitting rule — make the method easier to initialize and better at filling in…","keywords":["novel view synthesis","3D Gaussian Splatting","pointmap initialization","view-adaptive grouping","KNN-based densification","camera pose estimation","DUSt3R","adaptive density control"],"falsifier":"Take a scene where the two largest similarity gaps do not correspond to coherent viewpoint clusters (e.g., a textureless corridor with periodic appearance or a sequence with one large leap inside an otherwise dense orbit), run EasySplat's grouping with $k=2$, and compare the resulting ATE/RPE and final PSNR against a one-reference pairing of the same frames; if global alignment errors jump and render quality drops well below the oneref baseline, the grouping heuristic is the failing link.","tokens_in":9056,"feed_emoji":"🖼️","tokens_out":10031,"duration_ms":84325,"temperature":0.7,"pith_summary":"EasySplat claims that two targeted changes make 3D Gaussian Splatting (3DGS) both easier to initialize and better at filling missing detail. Instead of running a classic structure-from-motion pipeline, it cuts a dense image sequence into groups at the largest visual-similarity drops and pairs every non-reference frame in a group with the group's reference frame; those pairs feed a learned pointmap model whose pairwise reconstructions are globally aligned into camera poses and a point cloud. During training, it replaces the usual gradient-threshold densification with a rule that compares each Gaussian ellipsoid to the mean shape of its $k$ nearest neighbors and splits it when it is much larger. The paper reports that this combination reaches a mean PSNR of 33.800 and LPIPS of 0.056 on Tanks & Temples, ahead of the compared baselines, and also improves camera-pose accuracy over a COLMAP-free baseline on CO3DV2.","feed_headline":"Grouped pointmaps and KNN splits lift Gaussian splatting to 33.8 dB","feed_subtitle":"Replacing SfM with grouped pointmaps and neighbor-aware splitting beats prior baselines on Tanks & Temples.","key_machinery":"The two load-bearing mechanisms are the view-adaptive group initialization (Eqs. 1–4) and the KNN-based densification (Eq. 7). The grouping mechanism computes the cosine similarity of adjacent frames, forms the array $\\Delta$ of differences between consecutive similarity values, cuts the sequence where $\\Delta$ is largest ($k=2$ in the experiments), and matches every non-reference image in each group to that group's reference image; the resulting pairs are fed to a pretrained pointmap model (a model that predicts a 3D point map for an image pair), and Eq. (4) globally aligns the pairwise pointmaps into poses and a point cloud. The densification mechanism takes each Gaussian ellipsoid, finds its $n$ nearest neighbors (with $n=64$), averages their covariances to get $\\bar{\\Sigma}_{3d}$, and splits the ellipsoid when its own covariance is larger than this local average. Together they replace the two components the paper identifies as limiting: SfM-based sparse point initialization and gradient-threshold Adaptive Density Control.","core_discovery":"On the paper's own terms, the discovery is that the two most fragile parts of 3DGS — how the scene is initialized and when Gaussians are split — can both be fixed without adding heavy machinery. For initialization, it shows that a view-adaptive grouping rule (split the sequence at the $k$ largest gaps in adjacent-frame cosine similarity, then pair each non-reference view with that group's reference view) produces pointmap pairs whose global alignment is both memory-feasible on dense scenes and more accurate than a complete graph or a single-reference scheme. For optimization, it shows that splitting a Gaussian when its covariance exceeds the mean covariance of its $n$ nearest neighbors densifies regions where the initial point cloud is thin, recovering fine structure such as small objects. With those two components, EasySplat claims the best novel view synthesis results among the compared methods on Tanks & Temples and markedly better pose estimation than the compared COLMAP-free baseline on the harder CO3DV2 sequences.","pith_inferences":["One editorial extension the paper only gestures at is adaptive group counting: using a similarity threshold or a learned overlap estimate to choose $k$ instead of fixing $k=2$ could make the grouping rule transfer to sparse-view, exploratory, or orbit-like capture, and the grouping itself could be tested per group by measuring pairwise overlap before global alignment.","Because the KNN-split rule is parameter-light and orthogonal to gradient-based adaptive density control, it could plausibly be grafted onto anchor-based Gaussian methods or frequency-based densifiers; that is my inference, not a claim the paper tests.","A practical ceiling implied by the design is that the whole pipeline inherits the failure modes of the pretrained pointmap model: scenes with large appearance changes or low texture may still produce poor pairs even when grouping is perfect, so per-group pose accuracy should be monitored rather than assumed.","A testable extension from the paper's Table 1 is to replace the manual $k=2$ with a data-driven group count and check whether the known trade-off (more pairs improve pose but inflate GPU memory) can be optimized per scene."],"forward_implications":["Dense-view Gaussian splatting can be initialized without SfM: the grouping rule turns a costly complete pairing into a small set of pairs (1,142 instead of 159,600 on the Church scene), making learned pointmap priors practical for hundreds of input images.","If the KNN-split criterion is right, densification stops depending on gradient magnitude alone and becomes sensitive to local geometric scale, which is what lets it grow points in under-initialized regions and sharpen small objects.","Because camera poses are estimated alongside the point cloud, the method offers a route to pose-free novel view synthesis: pose errors on CO3DV2 improve over the compared non-COLMAP baseline while rendering quality also rises.","The reported Tanks & Temples numbers (mean PSNR 33.800, SSIM 0.953, LPIPS 0.056) put the full pipeline above ScaffoldGS and Mip-Splatting on the same benchmark, suggesting the two modifications compound rather than merely trade off."],"supporting_citations":[{"why":"Supplies the pairwise pointmap prior that replaces SfM initialization; its confidence-weighted alignment is used in Eq. (4).","marker":"[4]"},{"why":"Defines the 3D Gaussian Splatting representation and differentiable rasterizer that the proposed framework builds on.","marker":"[1]"},{"why":"The COLMAP SfM pipeline that the paper identifies as the fragile initialization being replaced.","marker":"[2]"},{"why":"CF-3DGS is the main COLMAP-free baseline and supplies the evaluation protocol on Tanks & Temples.","marker":"[3]"},{"why":"InstantSplat combines DUSt3R with 3DGS for sparse views and shows the complete-graph pairing that is too costly for dense views.","marker":"[5]"},{"why":"ScaffoldGS is the anchor-based SOTA baseline that EasySplat outperforms on the benchmark, and its densification approach is the point of comparison.","marker":"[7]"},{"why":"Mip-Splatting is a SOTA comparison addressing high-frequency artifacts and aliasing in Gaussian splatting.","marker":"[9]"},{"why":"The pretrained pointmap model variant used in the implementation to produce pairwise pointmaps for the grouping strategy.","marker":"[34]"},{"why":"Provides the Tanks & Temples benchmark scenes and protocol used for the main NVS comparison.","marker":"[28]"},{"why":"Provides the CO3DV2 scenes used to evaluate pose accuracy under large camera motions.","marker":"[29]"}],"fun_headline_variants":["EasySplat: view-adaptive grouping replaces SfM, KNN splits refine","No SfM needed: grouped pointmaps and neighbor-aware splitting for 3DGS","EasySplat: adaptive initialization and densification improves 3DGS","Grouped pointmaps and KNN splits outdo SfM and uniform splitting","View-adaptive pointmap grouping and KNN splits lift Gaussian splatting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two largest drops in adjacent-frame cosine similarity divide a dense sequence into groups in which every non-reference frame overlaps the group's reference frame well enough for pairwise pointmap alignment; the paper validates this heuristic with a single ablative comparison on the Church scene, so if the grouping misfires in other capture patterns the estimated poses and point cloud degrade before training can compensate.","fun_headline_variants_meta":{"raw":{"variants":["EasySplat: view-adaptive grouping replaces SfM, KNN splits refine","No SfM needed: grouped pointmaps and neighbor-aware splitting for 3DGS","EasySplat: adaptive initialization and densification improves 3DGS","Grouped pointmaps and KNN splits outdo SfM and uniform splitting","View-adaptive pointmap grouping and KNN splits lift Gaussian splatting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000299,"raw_usage":{"total_tokens":1735,"prompt_tokens":961,"completion_tokens":774,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":668}},"tokens_in":577,"tokens_out":774,"duration_ms":7466,"temperature":1.0,"reasoning_tokens":668,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:36:44.247318+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a scene where the two largest similarity gaps do not correspond to coherent viewpoint clusters (e.g., a textureless corridor with periodic appearance or a sequence with one large leap inside an otherwise dense orbit), run EasySplat's grouping with $k=2$, and compare the resulting ATE/RPE and final PSNR against a one-reference pairing of the same frames; if global alignment errors jump and render quality drops well below the oneref baseline, the grouping heuristic is the failing link.","supporting_citations":[{"cited_title":"Dust3r: Geometric 3d vision made easy,","cited_arxiv_id":null,"evidence_quote":"Supplies the pairwise pointmap prior that replaces SfM initialization; its confidence-weighted alignment is used in Eq. (4)."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering,","cited_arxiv_id":null,"evidence_quote":"Defines the 3D Gaussian Splatting representation and differentiable rasterizer that the proposed framework builds on."},{"cited_title":"Structure-from- motion revisited,","cited_arxiv_id":null,"evidence_quote":"The COLMAP SfM pipeline that the paper identifies as the fragile initialization being replaced."},{"cited_title":"Mip-splatting: Alias-free 3d gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Mip-Splatting is a SOTA comparison addressing high-frequency artifacts and aliasing in Gaussian splatting."},{"cited_title":"Tanks and temples: Benchmarking large-scale scene reconstruction,","cited_arxiv_id":null,"evidence_quote":"Provides the Tanks & Temples benchmark scenes and protocol used for the main NVS comparison."},{"cited_title":"Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruc- tion,","cited_arxiv_id":null,"evidence_quote":"Provides the CO3DV2 scenes used to evaluate pose accuracy under large camera motions."}],"review_version":1}