{"id":"adc03c0d-cb2a-4b35-a5f9-da9ede2fc19c","arxiv_id":"2507.08136","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RegGS aligns locally generated 3D Gaussian maps using a Sinkhorn-approximated mixture Wasserstein distance, improving pose estimation and novel view synthesis from sparse unposed views.","lead":"This paper presents RegGS, a system that stiches together 3D Gaussian scene fragments generated from sparse, unposed images into one coherent 3D model. It could make 3D reconstruction from a handful of casually shot photos more practical, with potential uses in AR, VR, and robotics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported metrics conflict internally: full RegGS 16-view RE10K ATE/PSNR differ between Table 3/1 and Table 4, leaving the central empirical claim without internally consistent support.","rationale":"The reader focused on the dependency on feed-forward Gaussian quality, which is a real limitation noted by the authors. However, a more immediate and checkable vulnerability is the internal inconsistency between the main results and the ablation table. Table 1/3 vs Table 4 report the same configuration (full method, 16 views, RE10K) but give very different PSNR and ATE values. This suggests either underexplained differences in protocol or unreliable numerical reporting. This does not by itself prove the central claim false, but it means the empirical support is not yet trustworthy. A conditional acceptance requiring the authors to release code and specify exact evaluation protocols is appropriate; we do not advocate rejecting the paper outright.","tokens_in":16653,"tokens_out":9320,"duration_ms":99808,"concrete_test":"Ask the authors to provide the exact evaluation protocol for Table 4 (test split, number of scenes, whether global refinement is applied, random seeds, and the sampling of 16 views) and to re-run the full RegGS pipeline under that protocol on the RE10K test set. Then verify that the resulting ATE and PSNR match Table 3/1 for the 16-view setting. If the numbers cannot be reconciled, the reported results are not reproducible and the claimed superiority over baselines is unverified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is an internal inconsistency in the reported experimental evidence. Table 3 reports RegGS ATE 0.041 at 16 views on RE10K, and Table 1 reports PSNR 28.663 at 16x. The ablation Table 4, which states 'the input is set to 16 views' on RE10K, reports the full model ('RegGS (Ours)') with ATE 0.098 and PSNR 23.09. These are the same configuration (full RegGS, 16 views, RE10K) yet the metrics differ by factors of ~2.4 and ~5.6 dB respectively. No explanation is given for the discrepancy (e.g., different test split, no global refinement, different number of scenes, seed). If the ablation is run under different conditions, the protocol is not reported; if it is run under the same conditions, the reported numbers are not reproducible. Because the central claim ('high-quality novel view synthesis' and 'precise pose estimation') rests on Tables 1 and 3, and because the only internal check (Table 4) does not reproduce them, the evidence is not yet internally consistent. This is a concrete, checkable problem that should be resolved before the central claim is accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RegGS, a framework for reconstructing 3D Gaussian scenes from unposed sparse views by incrementally registering local 3D Gaussian maps produced by a pre-trained feed-forward model (NoPoSplat). The alignment is performed in Sim(3) space using an entropy-regularized Sinkhorn approximation of the Mixture 2-Wasserstein (MW2) distance between Gaussian mixture models, combined with photometric and depth-consistency losses in a joint differentiable registration module, followed by global refinement. The paper reports state-of-the-art novel-view synthesis and pose-estimation results on RE10K and ACID, with an ablation study of the loss components and the registration module. The appendix provides a claimed gradient-consistency proof for the entropy-regularized MW2 distance and a complexity analysis of the Sinkhorn algorithm.","tokens_in":16989,"tokens_out":5587,"duration_ms":54642,"significance":"If the reported results are correct, RegGS is a practically valuable contribution: it extends feed-forward Gaussian predictors beyond their fixed input-size limitation, uses a differentiable optimal-transport metric for Gaussian-mixture alignment, and shows large improvements over unposed optimization-based baselines on two benchmarks. The paper includes useful qualitative comparisons, runtime/memory measurements, and an honest discussion of failure modes. However, the central empirical evidence contains an unresolved internal inconsistency between the main results and the ablation tables, and the theoretical justification for the gradient-consistency claim is not rigorous. These issues need to be addressed before the main claims can be fully accepted.","major_comments":[{"comment":"There is an internal inconsistency in the reported 16-view RE10K results. Table 3 reports RegGS with ATE 0.041 at 16 views, and Table 1 reports PSNR 28.663 for the same setting. However, Table 4, which states 'the input is set to 16 views' on RE10K and evaluates the full model ('RegGS (Ours)'), reports ATE 0.098 and PSNR 23.09. Unless the ablation is run under a different protocol (no global refinement, different split, different seed, or a subset of scenes), these numbers are irreconcilable. Please state the exact protocol for the ablation experiments and explain the discrepancy. Because the central claims of precise pose estimation and high-quality novel-view synthesis rest on Tables 1 and 3, this inconsistency undermines the empirical support and must be resolved.","section":"Sec. 4.3, Tables 1, 3, and 4"},{"comment":"The gradient-consistency proof is not a valid proof. Gamma-convergence of the functionals in Eq. (20) concerns convergence of minimizers and does not imply convergence of gradients; Eq. (22) invokes an implicit-function/envelope-theorem argument without verifying the required differentiability and regularity conditions for the entropic optimal transport plan pi*_ik(epsilon); and Eq. (23) asserts convergence of the regularized plans to an optimal plan, which is not established by Gamma-convergence. Consequently, the claimed limit in Eq. (25), lim_{epsilon->0} grad W2,epsilon in subdifferential W2, is unsupported. Since this claim is used to justify the decreasing-epsilon optimization schedule, the proof should either be made rigorous or the theoretical statement should be substantially softened.","section":"Supplementary Sec. 6, Eqs. (19)-(25)"},{"comment":"The MW2 distance is used both as the alignment objective in Eq. (13) and as the '3DGS registration evaluation' metric in Sec. 4.1 and in the ablation tables. This creates a circularity: the MW2 column in the ablations largely measures how well the trained loss is minimized, not an independent geometric quality. The ATE and NVS metrics are independent and mitigate this, but the registration-quality claim should be validated with an independent geometric measure (for example, Chamfer distance between point clouds or rendered depth maps), or the MW2 column should be presented explicitly as a loss-curve indicator rather than as an evaluation metric.","section":"Sec. 4.1 and Tables 4-5"},{"comment":"The paper does not report several free parameters that are essential for reproducibility: the loss weights lambda1, lambda2, lambda3 in Eq. (13), the entropy regularization parameter epsilon, the number of Sinkhorn iterations T, the optimizer settings (learning rate, number of iterations), and the specifics of scale normalization. An experimental paper whose central claim is a new optimization-based pipeline should provide these values in the main text or supplementary material. Please add a complete implementation-details subsection.","section":"Sec. 4.1, Implementation Details"}],"minor_comments":[{"comment":"The first paragraph says the main map and submap contain 'M and N Gaussian gradients'; this should be 'Gaussian components' or 'Gaussians'.","section":"Supplementary Sec. 7"},{"comment":"The phrase 'the two sets of Gaussian spheres are not in one-to-one correspondence' should be rephrased to 'Gaussian components' for consistency with the rest of the paper.","section":"Sec. 3.2"},{"comment":"The baseline name is written as 'MARSt3R [24]' in the text but as 'MASt3R' in Tables 1 and 2; please make the notation consistent.","section":"Sec. 4.1, Baselines"},{"comment":"The text says 'As shown in Tab. 7, RegGS maintains controlled memory usage', but the corresponding item is presented as Figure 7; the cross-reference should be corrected.","section":"Supplementary Sec. 8"},{"comment":"The limitation discussion would be clearer if it explicitly referenced the supplementary failure example, since the main text's reference to reflective-surface failures is otherwise not tied to a figure.","section":"Sec. 4.4 and Supplementary Fig. 9"}],"recommendation":"major_revision","confidential_remarks":"The main blocker is the unresolved discrepancy between Tables 1/3 and Table 4 for the 16-view RE10K setting. The authors should be asked to clarify whether the ablation uses a different protocol, and if so, to report it. The supplementary proof of gradient consistency appears to be a placeholder and should be either fixed or removed. The paper is within scope and the core idea is promising, so I would not reject it, but the current version is not yet internally consistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is a sensible integration: use a feed-forward 3DGS model (NoPoSplat) to generate local Gaussian maps, then align them in Sim(3) with a differentiable MW2 distance computed via entropy-regularized Sinkhorn, plus photometric and depth losses. That genuinely extends feed-forward reconstruction from strictly two views to 8, 16, or 32 views without poses. The engineering details are thoughtful—quaternion parameterization, log-space Sinkhorn, Cholesky-based covariance terms, scale normalization—and the paper is honest about upstream-model failures and visual artifacts in the supplementary. The soft spots are real, and one is serious. The ablation table (Table 4) reports the full model (RegGS, Ours) on RE10K at 16 views with ATE 0.098 and PSNR 23.09. The main tables report the same configuration with ATE 0.041 (Table 3) and PSNR 28.663 (Table 1). Those are very different numbers. The paper offers no explanation—no mention of a different split, no note about omitting global refinement, no seed variance. Since the headline claims rest on Tables 1 and 3, this internal inconsistency undercuts the empirical story. The ablation's relative trends (each loss matters) are plausible, but the absolute numbers are suspect until the discrepancy is resolved. Also worth flagging: MW2 is used both as a training loss and as the registration evaluation metric (Sec 4.1), which is circular for the registration-specific claim, though the NVS and pose metrics are independent and do provide separate grounding. The supplementary's gradient-consistency proof is asserted rather than rigorous—Γ-convergence of function values does not imply gradient convergence, and the Clarke subgradient argument is sketched, not proved. That is a fixable issue. Hyperparameters (λ1..3, ε, Sinkhorn iterations, optimization schedule) are not reported, and no code is released, so reproduction is currently impossible. Who is this for: people working on pose-free sparse-view reconstruction, especially those building on feed-forward Gaussian models. The method is a reasonable next step and deserves a serious referee. I would send it to review, but ask for a corrected and consistent experimental table, full hyperparameters, and either a proper gradient-convergence argument or a softened claim.","headline":"A plausible engineering integration for registering feed-forward 3D Gaussians beyond two views, but the main experimental tables contradict the ablation table on the same 16-view RE10K configuration, so the reported gains are not yet internally consistent.","tokens_in":737,"tokens_out":810,"would_cite":false,"duration_ms":38720,"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":"RegGS reconstructs scenes from sparse unposed images by registering feed-forward 3D Gaussian maps with optimal transport, reporting precise poses and high-quality novel views.","keywords":["3D Gaussian Splatting","unposed sparse view reconstruction","Gaussian mixture model registration","Mixture 2-Wasserstein distance","Sinkhorn optimal transport","Sim(3) alignment","camera pose estimation","novel view synthesis"],"falsifier":"Take a scene dominated by a reflective or textureless surface, the failure mode shown in Fig. 9: if the upstream model misplaces the depth of those Gaussians, the MW2, photometric, and depth losses should fail to register the sub-map, and ATE and PSNR should degrade sharply even though the same method succeeds on diffuse scenes. A quantitative version is to compare this ATE gap to the upstream depth error on the same surfaces.","tokens_in":1536,"feed_emoji":"🧊","tokens_out":1613,"duration_ms":82565,"temperature":0.7,"pith_summary":"RegGS tackles the problem of reconstructing a 3D scene from a handful of images with no known camera poses. The paper's claim is that instead of optimising a global scene directly, you can let a pretrained feed-forward network predict local 3D Gaussian maps for each input and then register those maps into one consistent representation. The registration is driven by a Mixture 2-Wasserstein distance between Gaussian mixture models, computed efficiently with an entropy-regularized Sinkhorn algorithm, plus photometric and depth losses. Concretely, on RE10K with 16 views it reports PSNR 28.663 and an ATE of 0.041, better than the unposed baselines it compares against.","feed_headline":"One registration step gives precise poses from unposed sparse views","feed_subtitle":"Aligning feed-forward 3D Gaussians with an optimal-transport metric beats unposed baselines on RE10K and ACID.","key_machinery":"The load-bearing object is the Mixture 2-Wasserstein ($\\mathrm{MW}_2$) distance between two Gaussian mixture models, computed with an entropy-regularized Sinkhorn algorithm. This differentiable metric measures how far the transformed sub-map's Gaussians are from the main map's Gaussians, including both means and covariances, and it is what lets the registration solve for a $\\mathrm{Sim}(3)$ transformation. The joint registration module then combines this structural term with rendered photometric and depth losses, refining the coarse optimal-transport alignment into precise geometry.","core_discovery":"The paper claims that local 3D Gaussian maps, each generated from a small number of images by a pretrained feed-forward model, can be fused into a single globally consistent 3D Gaussian scene by solving for the similarity transformation (rotation, translation, and uniform scale) in $\\mathrm{Sim}(3)$ that minimizes an optimal-transport distance between the maps treated as Gaussian mixture models. It further claims that combining this structural distance with photometric and depth consistency losses in a differentiable joint registration module yields accurate camera poses and high-quality novel views from unposed sparse images, demonstrated on RE10K and ACID. The method is incremental: two images seed a main map, each new view contributes a sub-map that is aligned and merged, and a final global refinement polishes the result.","pith_inferences":["The same optimal-transport registration could be applied to align sub-maps from other feed-forward 3D predictors or even from NeRF-style reconstructions whose output has been converted to Gaussian components.","A practical next step is to warm-start the Sinkhorn transport plan from the previous frame during incremental scanning, which would lower the reported runtime bottleneck without changing the method.","The documented failure on reflective surfaces suggests that training the upstream generator with explicit depth supervision on specular geometry would likely reduce registration failures more than improving the registration module itself.","The large pose-accuracy gains at 8 views point toward sequential SLAM and embodied-agent settings, where poses must be recovered online from a stream of sparse views."],"forward_implications":["The same registered map directly yields a camera pose for every input view, removing the need for a separate structure-from-motion stage.","The approach is not capped at the two-view limit of feed-forward models: as more views arrive, each new local map is merged into the growing global map and a final refinement polishes the whole scene.","Because the structural MW2 term ignores colour, the alignment stage is insensitive to lighting changes across views, while the photometric and depth terms then refine local geometry.","Removing any one of the three losses sharply degrades both pose and rendering quality, so the joint registration design is the load-bearing contribution."],"supporting_citations":[{"why":"Supplies the pretrained feed-forward model that generates the local 3D Gaussian sub-maps to be registered.","marker":"[43]"},{"why":"Defines the Mixture 2-Wasserstein distance between Gaussian mixture models used as the structural alignment metric.","marker":"[17]"},{"why":"Provides the entropy-regularized Sinkhorn iterations that make the MW2 distance efficiently computable and differentiable.","marker":"[15]"},{"why":"Supplies the differentiable 3D Gaussian splatting rendering pipeline behind the photometric and depth losses.","marker":"[23]"},{"why":"Optimization-based unposed baseline whose sparse-view results RegGS must beat.","marker":"[20]"},{"why":"RE10K dataset used for the novel-view synthesis and pose estimation evaluations.","marker":"[46]"},{"why":"ACID dataset used to test generalization to drone-captured aerial scenes.","marker":"[26]"}],"fun_headline_variants":["Optimal transport aligns 3D Gaussians for unposed sparse views","From sparse unposed images to precise poses via Gaussian fusion","Register unposed sparse views into one scene with Gaussian alignment","RegGS: Unposed sparse views meet Gaussian registration","One alignment step yields precise poses for unposed sparse views"],"cache_read_input_tokens":19584,"weakest_assumption_plain":"The pipeline rests on the pretrained feed-forward model producing local 3D Gaussians that are structurally reliable enough that a single similarity transform can align each sub-map with the main map; if any sub-map's geometry is poor, the registration losses cannot repair it.","fun_headline_variants_meta":{"raw":{"variants":["Optimal transport aligns 3D Gaussians for unposed sparse views","From sparse unposed images to precise poses via Gaussian fusion","Register unposed sparse views into one scene with Gaussian alignment","RegGS: Unposed sparse views meet Gaussian registration","One alignment step yields precise poses for unposed sparse views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000677,"raw_usage":{"total_tokens":3086,"prompt_tokens":957,"completion_tokens":2129,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2046}},"tokens_in":573,"tokens_out":2129,"duration_ms":15710,"temperature":1.0,"reasoning_tokens":2046,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:27:25.873961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a scene dominated by a reflective or textureless surface, the failure mode shown in Fig. 9: if the upstream model misplaces the depth of those Gaussians, the MW2, photometric, and depth losses should fail to register the sub-map, and ATE and PSNR should degrade sharply even though the same method succeeds on diffuse scenes. A quantitative version is to compare this ATE gap to the upstream depth error on the same surfaces.","supporting_citations":[{"cited_title":"A wasserstein-type dis- tance in the space of gaussian mixture models.SIAM Journal on Imaging Sciences, 13(2):936–970, 2020","cited_arxiv_id":null,"evidence_quote":"Defines the Mixture 2-Wasserstein distance between Gaussian mixture models used as the structural alignment metric."},{"cited_title":"Sinkhorn distances: lightspeed computation of optimal transport","cited_arxiv_id":null,"evidence_quote":"Provides the entropy-regularized Sinkhorn iterations that make the MW2 distance efficiently computable and differentiable."},{"cited_title":"Stereo magnification: Learning view syn- thesis using multiplane images","cited_arxiv_id":null,"evidence_quote":"RE10K dataset used for the novel-view synthesis and pose estimation evaluations."},{"cited_title":"Infinite na- ture: Perpetual view generation of natural scenes from a sin- gle image","cited_arxiv_id":null,"evidence_quote":"ACID dataset used to test generalization to drone-captured aerial scenes."}],"review_version":1}