{"id":"1df129ba-543a-48fb-b565-76b7c53dad8d","arxiv_id":"2507.18541","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An unposed 3DGS system that aligns per-group MVS reconstructions via probabilistic Procrustes mapping and jointly optimizes Gaussians and poses, reporting state-of-the-art results on Waymo and KITTI.","lead":"This paper proposes a pipeline that reconstructs 3D Gaussian Splatting scenes from hundreds of unpositioned outdoor photos by dividing them into groups, aligning each group's point cloud with a probabilistic Procrustes mapping, and then jointly refining the scene and camera poses. It reports strong results on Waymo and KITTI, but the evaluation uses all images for both fitting and scoring, which weakens the rendering-quality claims.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reconstruction SOTA is unsupported because PSNR/SSIM/LPIPS are computed on the same training views used for joint optimization; a held-out split is required, and Table 2 'Ours' conflicts with Table 1 (Waymo ATE 0.56 vs 1.41).","rationale":"The paper's central claim is threefold: the probabilistic Procrustes mapping aligns submaps, the joint optimization refines poses and Gaussians, and the full system achieves state-of-the-art unposed 3DGS reconstruction on Waymo and KITTI. The third part is the headline claim. For it to hold, the evaluation must actually measure reconstruction quality. Section 4.1 explicitly states that all images are used for evaluation, and Table 1's PSNR/SSIM/LPIPS are computed on the same images used to train the 3DGS scene and optimize poses through Eq. (14). With joint pose and Gaussian optimization, the model can fit training views well regardless of whether the underlying geometry is accurate; high training-view metrics are therefore weak evidence for a reconstruction claim. The relative comparison to baselines is internally consistent because all methods are evaluated the same way, but that does not justify the 'new state of the art' phrasing for unposed reconstruction. The numerical conflict between Table 1 and Table 2 for 'Ours' on Waymo (ATE 1.41 vs 0.56, PSNR 31.53 vs 32.72) compounds this by making even the training-view numbers non-reproducible from the paper. The reader's weakest assumption regarding dependency on VGGT and the single overlapping frame is a legitimate robustness concern, and the paper's Limitations section acknowledges the MVS dependence. However, the evaluation-protocol flaw is more load-bearing because it puts the primary quantitative support for the reconstruction half of the central claim in question. A held-out split and a clarification of the conflicting tables would settle the issue; if the results are maintained under that test, the central claim is substantially better supported. Since the reader's verdict is already CONDITIONAL with moderate confidence, my analysis does not move the verdict.","tokens_in":17010,"tokens_out":5662,"duration_ms":59760,"concrete_test":"Re-run the Waymo and KITTI experiments with a strict train/test split, for example optimizing on 100 frames per sequence and evaluating PSNR/SSIM/LPIPS on the remaining 100 held-out frames, and report the same metrics for all baselines. If the held-out PSNR drops substantially or the margin over DUSt3R*/MASt3R* narrows, the reconstruction SOTA claim is unsupported. Separately, the authors should clarify which experimental configuration produced Table 1 versus Table 2 and reconcile the differing Waymo 'Ours' ATE/PSNR values, ideally with per-scene results and error bars.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the method achieves state-of-the-art unposed 3DGS reconstruction from hundreds of outdoor images. For that claim to hold, the reported reconstruction metrics must measure reconstruction quality, not training-set fitting. Section 4.1 states 'All images are used for evaluation,' and the PSNR/SSIM/LPIPS values in Table 1 are computed on the same frames that Eq. (14) optimizes against with L1 and SSIM losses. Since the joint optimization module also updates per-view camera poses, 3DGS can overfit these training views; high training-view metrics do not demonstrate that the inferred scene geometry or appearance generalizes to unseen viewpoints. This directly undercuts the 'accurate reconstruction' and 'new state of the art' conclusions in the abstract and Section 4.2. The evaluation protocol is the same for all baselines, so the relative ranking is a fair comparison of training-view fitting, but it does not support the reconstruction claim. Compounding this, the paper reports inconsistent numbers for the same method and dataset: Table 1 lists 'Ours' on Waymo with ATE 1.41 and PSNR 31.53, while Table 2 lists 'Ours' with ATE 0.56 and PSNR 32.72, with no explanation of which configuration produced each table. Without a held-out evaluation and a resolution of this discrepancy, the headline quantitative evidence for SOTA reconstruction is not reliable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unposed 3D Gaussian Splatting (3DGS) reconstruction framework for hundreds of outdoor images. The method partitions images into overlapping subsets, runs a pretrained MVS model (VGGT) on each subset to obtain local point clouds and camera poses, then aligns all submaps into a global coordinate frame via a probabilistic Procrustes mapping. The alignment is initialized with a closed-form Kabsch–Umeyama Sim(3) solution and refined with entropy-regularized soft correspondences plus a dustbin mechanism for outlier rejection. A final stage jointly optimizes 3D Gaussians and camera poses through differentiable rendering with analytic Jacobians. Experiments on Waymo and KITTI report state-of-the-art pose accuracy (ATE) and image reconstruction metrics (PSNR, SSIM, LPIPS) compared with COLMAP-based, feed-forward, and unposed 3DGS baselines. The central claim is that the method removes the SfM bottleneck for large-scale unposed 3DGS while achieving accurate global alignment in minutes.","tokens_in":17384,"tokens_out":4325,"duration_ms":47523,"significance":"If the empirical claims hold under a proper evaluation protocol, the paper addresses a real scalability bottleneck: unposed 3DGS from hundreds of images with pretrained MVS models and divide-and-conquer alignment. The probabilistic Procrustes formulation and the analytic pose gradients are technically sensible, and the reported efficiency (minutes for global alignment) is potentially useful. The derivation is not circular in itself; it optimizes a well-defined objective on external MVS outputs. However, the current evaluation is training-view fitting rather than novel-view synthesis, and the headline numbers are internally inconsistent. Since the central claims depend on these numbers, the significance can only be assessed after the evaluation issues are resolved. The analytic Jacobian derivations in the appendices are a concrete strength that would help reproducibility if the full method were specified.","major_comments":[{"comment":"Section 4.1 states 'All images are used for evaluation,' and Eq. (14) optimizes L1 and SSIM over exactly these images; consequently the PSNR/SSIM/LPIPS values in Table 1 are training-set fit scores, not evidence of reconstruction or novel-view synthesis quality. The abstract and Section 4.2 claim 'accurate reconstruction' and 'new state of the art' on the strength of these numbers; please add a held-out split (or at least a train/test separation) and report both fitting and generalization metrics, or explicitly restate the claim as training-view fitting quality.","section":"§4.1, Eq. (14)"},{"comment":"The Waymo row for 'Ours' differs across tables: Table 1 reports ATE 1.41, PSNR 31.53, SSIM 0.915, LPIPS 0.245, while Table 2 reports ATE 0.56, PSNR 32.72, SSIM 0.935, LPIPS 0.211. No explanation is given for the discrepancy, and the ablation row 'w/o Joint Opt.' (ATE 0.68) is already better than the Table 1 headline (ATE 1.41); the paper must specify the configuration for each table and reconcile these numbers, otherwise readers cannot determine which result is the claimed state of the art.","section":"Tables 1 and 2"},{"comment":"The dustbin mechanism is claimed as a core contribution but is underspecified: η and b_dustbin = δ are introduced in the text, yet they do not appear in objective (6) or in the weight update (7), and the relationship between the dustbin marginal and the constraint Σ_l γ_l = 1 is not given. Please provide the complete optimization problem including the dustbin variable, its marginal constraint, and how δ is set.","section":"§3.2.2, Eq. (6)"},{"comment":"Correspondences are built from a single shared frame (K=1) between adjacent 60-image groups, so the Procrustes alignment in Eq. (2) is only as good as the MVS depths and poses of that one overlapping frame; the paper acknowledges dependence on MVS quality in §4.4 but reports no sensitivity analysis for K, group size, or correspondence confidence. A robustness study varying K and group size is needed to support the claim of scalable and robust alignment for hundreds of images.","section":"§4.1, Eq. (1), Limitations (§4.4)"}],"minor_comments":[{"comment":"The paper does not report error bars or statistical significance; since the claims include small margins over baselines (e.g., PSNR 31.53 vs 30.17 for COLMAP+SPSG on Waymo), per-scene variance and run-to-run variance should be reported.","section":"Tables 1 and 2"},{"comment":"The entropy regularization coefficient ε in Eq. (6) and the loss weight α in Eq. (14) are not given numerical values or ranges in the implementation details; please add them.","section":"§3.2.2 and §3.3"},{"comment":"The caption of Figure 2 states InstantSplat is trained on only 80 images, but Table 1 does not state this configuration; please clarify which rows correspond to 80 vs 200 images.","section":"§4.1, Figure 2"},{"comment":"The text alternates between 'Tab. 1' and 'Table 1', and between 'Mast3R' and 'MASt3R'; please standardize the notation and reference style.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is an arXiv preprint under review, and the NeurIPS checklist claims that all experimental details are provided, but several central hyperparameters and the exact evaluation protocol are missing from the main text. The internal inconsistency between Table 1 and Table 2 for the same method is the most serious issue and should be resolved before any accept decision. The authors should also be asked whether the reported runtimes include preprocessing, correspondence extraction, and the joint optimization, since the abstract's 'within minutes' claim is an important part of the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick read on arXiv:2507.18541 (Cheng et al., \"Unposed 3DGS Reconstruction with Probabilistic Procrustes Mapping\").\n\nThe core idea is genuinely new and worth attention: partition hundreds of unposed images into overlapping subsets, let VGGT produce local point clouds and poses, then align those submaps with a probabilistic Procrustes solver that includes a soft dustbin to reject bad correspondences. The Kabsch–Umeyama closed-form init followed by an entropy-regularized probabilistic refinement is sensible, and the analytic quaternion Jacobian for pose-gradient backprop through 3DGS is a nice engineering detail that avoids AD overhead. Speed is a real selling point: minutes to align tens of millions of points, and the qualitative figures show markedly better submap consistency than ICP or COLMAP-relative-pose alternatives. That part of the paper stands.\n\nThe soft spot is the evaluation, and it is load-bearing for the \"state of the art\" claim. Section 4.1 says \"All images are used for evaluation,\" and Eq. (14) optimizes the very same L1+SSIM objective against those images. So the PSNR/SSIM/LPIPS numbers in Table 1 measure how well the Gaussians fit the training views, not whether the reconstruction generalizes to held-out viewpoints. Every baseline suffers the same protocol, so the relative ranking is a fair comparison of training-set fitting, but it cannot support the abstract's \"accurate reconstruction\" and \"new state of the art\" conclusions. The fix is standard: hold out a subset of frames per scene for testing and report those numbers.\n\nAlso, Table 1 lists \"Ours\" on Waymo at ATE 1.41 and PSNR 31.53, while Table 2 lists \"Ours\" at ATE 0.56 and PSNR 32.72. No explanation is given for which configuration produced which table. That is exactly the kind of inconsistency a referee will catch, and it needs a clear footnote or corrected numbers. Minor but related: several hyperparameters (dustbin capacity, entropy coefficient, anchor retention ratios) are given only in prose or not at all, and there are no error bars — the checklist honestly says \"No\" on statistical significance, which is at least candid.\n\nOn the positive side, the derivation of the probabilistic mapping is coherent, the limitations section honestly acknowledges dependence on MVS quality and drift on long sequences, and the related work is well placed. The internal math is not circular; the circularity is entirely in the evaluation protocol.\n\nBottom line: this is a promising method with a real contribution in submap fusion, but the quantitative evidence, as reported, does not yet justify the SOTA headline. It deserves a serious referee who will require a proper train/test split and a resolution of the Table 1/2 discrepancy. I'd send it to review with major revision expectations, not desk reject.","headline":"Solid submap-alignment idea with a clean analytic-jacobian touch, but the headline SOTA claim is undercut by an all-training-image evaluation and an unexplained Table 1/2 ATE discrepancy.","tokens_in":17911,"tokens_out":3216,"would_cite":true,"duration_ms":30507,"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":"Unposed 3DGS reconstruction from hundreds of outdoor images becomes practical via submap splitting, probabilistic Procrustes alignment, and joint pose-scene refinement.","keywords":["3D Gaussian Splatting","unposed reconstruction","probabilistic Procrustes","multi-view stereo","camera pose estimation","global alignment","novel view synthesis","divide-and-conquer"],"falsifier":"Run the pipeline on a sequence where the single overlap frame between two subsets is nearly degenerate (low parallax, heavy occlusion, or motion blur) and check whether the ATE at the submap boundary jumps to the level of the ICP baseline; alternatively, swap the pretrained MVS backbone for a weaker model and measure whether the final pose error degrades by more than the joint optimization can absorb.","tokens_in":16828,"feed_emoji":"🧩","tokens_out":7070,"duration_ms":64137,"temperature":0.7,"pith_summary":"The paper claims a way to run 3D Gaussian Splatting reconstruction from hundreds of unposed outdoor photos, without the usual Structure-from-Motion preprocessing. It splits the images into overlapping groups, lets a pretrained multi-view stereo network build a local point cloud and poses per group, and then registers those submaps into one global scene. The registration step is recast as a probabilistic Procrustes problem: a closed-form Sim(3) solution (Kabsch-Umeyama) seeds an entropy-regularized refinement with a dustbin channel that rejects unreliable correspondences. After global alignment, the paper jointly optimizes the Gaussians and the camera poses through differentiable rendering, using an analytic quaternion Jacobian for pose gradients. On Waymo and KITTI sequences of 200 images, it reports lower absolute trajectory error and better rendering quality than the tested baselines, with the alignment itself finishing in minutes.","feed_headline":"Submap alignment makes unposed 3DGS scale to hundreds of images","feed_subtitle":"A probabilistic Procrustes step merges local reconstructions, beating SfM pipelines on pose accuracy and rendering quality.","key_machinery":"The central mechanism is probabilistic Procrustes mapping (PPM). It takes pixel-level correspondences between an overlapping frame in two adjacent submaps - points that project to the same pixel under both local poses - and treats the submap-to-submap similarity transformation as a Procrustes problem with learnable soft correspondence weights. A closed-form Kabsch-Umeyama step provides the initial scale, rotation, and translation; then the objective L = sum_l gamma_l ||sR p_l + t - q_l||^2 + epsilon sum_l gamma_l ln gamma_l is minimized alternately over the weights (an entropy-regularized update, with a virtual dustbin point carrying a fixed marginal weight to reject outliers) and over the transformation (gradient descent on scale, translation, and a unit-quaternion rotation). A second mechanism is the joint 3DGS-pose optimization: Gaussians are initialized from confidence-aware anchor points, and camera poses receive analytic gradients through the renderer via a closed-form quaternion-point Jacobian and projected gradient normalization to keep the quaternion unit-length.","core_discovery":"On its own terms, the paper's discovery is that the bottleneck to unposed 3DGS at outdoor scale is not the Gaussian optimization but the global registration of many local reconstructions, and that this registration can be solved cleanly at the pixel level. Because adjacent image subsets share at least one frame, the corresponding 3D points in the overlapping frame define a correspondence set between submaps. Mapping those correspondences is formulated as a weighted Procrustes problem over Sim(3), solved first in closed form and then refined by probabilistic coupling with a soft dustbin that absorbs a controlled fraction of outliers. Iterating this over all adjacent pairs yields a globally consistent point cloud and camera trajectory. The paper then shows that initializing Gaussians from confidence-pruned, downsampled anchor points and refining poses through differentiable 3DGS rendering removes the residual scale noise and pose drift, achieving the reported state-of-the-art results on Waymo and KITTI.","pith_inferences":["The same divide-and-conquer alignment could be applied to submaps from any feed-forward MVS model, not just the one used here, so the framework's accuracy should scale with improvements in the backbone; a natural test is swapping in a different pretrained model and measuring whether ATE and PSNR improve in step.","Setting the overlap to a single frame means the entire alignment of a long sequence rests on the reliability of one shared frame per boundary; increasing the overlap to two or three frames would add redundancy at modest cost and might reduce drift on longer sequences, an extension the paper does not evaluate.","The probabilistic Procrustes formulation could transfer to other multi-submap fusion problems, such as simultaneous localization and mapping or large-scale point cloud assembly, where scale ambiguity and outlier correspondences are the dominant failure modes.","The reported ATE numbers are averages over scene groups; publishing per-sequence spreads would clarify whether the method's margin is consistent across scenes or driven by a few easy ones."],"forward_implications":["Unposed 3DGS reconstruction no longer requires a full COLMAP pass; the alignment step replaces hours of SfM with minutes of closed-form plus iterative refinement.","The method's pose accuracy (ATE 1.41 m on Waymo, 1.64 m on KITTI) and rendering quality (31.53/24.83 PSNR) beat the compared SfM-based and feed-forward baselines on 200-image outdoor sequences.","Because each subset is processed independently, the approach sidesteps the memory blow-up that limits feed-forward MVS and Splatting models to a dozen or so views.","The dustbin mechanism makes the alignment robust to a controlled fraction of unreliable correspondences, capped by a user-set parameter (20% in the experiments).","The ablations show that disabling either the probabilistic Procrustes module or the joint 3DGS-pose refinement degrades both pose accuracy and rendering quality, indicating that both components carry the final result."],"supporting_citations":[{"why":"Pretrained MVS model VGGT supplies the local submap point clouds and poses that the framework aligns and refines.","marker":"[Wang et al., 2025a]"},{"why":"Provides the closed-form least-squares Sim(3) solution (Kabsch-Umeyama) that initializes the probabilistic Procrustes refinement.","marker":"[Umeyama, 1991]"},{"why":"Gives the algebraic justification for the Kabsch-Umeyama closed-form solution used in the initial alignment step.","marker":"[Lawrence et al., 2019]"},{"why":"3D Gaussian Splatting is the scene representation and differentiable renderer used for joint pose-scene optimization.","marker":"[Kerbl et al., 2023]"},{"why":"DUSt3R is the feed-forward pointmap baseline against which the MVS approach is compared in experiments.","marker":"[Wang et al., 2024b]"},{"why":"COLMAP-based SfM is the traditional pipeline the method aims to replace, and its SPSG variant is the main SfM baseline.","marker":"[Schonberger and Frahm, 2016]"},{"why":"Waymo Open Dataset provides one of the two outdoor evaluation suites (9 scene groups, 200 images each).","marker":"[Sun et al., 2020]"},{"why":"KITTI provides the second outdoor evaluation suite (8 scene groups).","marker":"[Geiger et al., 2013]"}],"fun_headline_variants":["Closed-form Procrustes alignment scales unposed 3DGS to hundreds of images","Submap merging with probabilistic Procrustes beats SfM for unposed 3DGS","Unposed 3DGS without SfM: probabilistic Procrustes merges submaps","Joint pose and Gaussian optimization from submap-aligned point clouds","Probabilistic Procrustes mapping enables global alignment in minutes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes the pretrained multi-view stereo network produces locally consistent point clouds and poses, and that the single overlapping frame between adjacent subsets yields reliable pixel-level correspondences; if the local depth has scale bias or the overlap view lacks parallax, the Procrustes alignment inherits that error.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form Procrustes alignment scales unposed 3DGS to hundreds of images","Submap merging with probabilistic Procrustes beats SfM for unposed 3DGS","Unposed 3DGS without SfM: probabilistic Procrustes merges submaps","Joint pose and Gaussian optimization from submap-aligned point clouds","Probabilistic Procrustes mapping enables global alignment in minutes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1614,"prompt_tokens":1018,"completion_tokens":596,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":487}},"tokens_in":634,"tokens_out":596,"duration_ms":5576,"temperature":1.0,"reasoning_tokens":487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:32:22.471897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a sequence where the single overlap frame between two subsets is nearly degenerate (low parallax, heavy occlusion, or motion blur) and check whether the ATE at the submap boundary jumps to the level of the ICP baseline; alternatively, swap the pretrained MVS backbone for a weaker model and measure whether the final pose error degrades by more than the joint optimization can absorb.","supporting_citations":[],"review_version":1}