{"id":"aebb34b3-4f9b-4438-8bf3-594dcec44a83","arxiv_id":"2506.21629","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A depth-plus-ICP initialization and voxel-based densification let 3D Gaussian Splatting work without SfM, improving pose and rendering on large outdoor scenes.","lead":"The paper describes a way to reconstruct large outdoor 3D scenes from ordinary video without first computing camera positions through the slow structure-from-motion step. The method uses depth estimates plus an iterative point-alignment algorithm, and reports sharper renderings and more accurate trajectories than two existing pose-free approaches.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on unverified per-frame metric-depth scale consistency from Metric3DV2; if scale drifts over long KITTI-360 drives, G-ICP initialization and voxel densification inherit the error, and the reported sub-millimeter ATE values are implausible enough to warrant a direct…","rationale":"The reader's weakest assumption and my stress-test converge on the same load-bearing point: the entire pose-estimation cascade depends on Metric3DV2 depth being accurate and scale-consistent across long outdoor sequences, and that assumption is never directly tested. I agree with the reader's CONDITIONAL verdict, and I did not find a different concern that would force a change. I added one sharpening: Table 3's ATE values near 0.001 are so low that they are hard to reconcile with sequential monocular estimation over 100-m trajectories, so the pose-evaluation protocol itself needs scrutiny. The proposed concrete test is the smallest experiment that would settle whether depth-scale drift matters: normalize or replace the depth maps with KITTI-360 LiDAR ground truth only in the G-ICP/lifting stage and measure the downstream effect. Because the paper's ablations, code release, and internally consistent tables provide some support, the correct disposition remains conditional rather than reject; the claim is plausible but not fully supported until depth-scale robustness and pose-error definitions are clarified.","tokens_in":7865,"tokens_out":10367,"duration_ms":120857,"concrete_test":"Compute per-frame scale factor s_t = median(D_pred / D_lidar) on the five KITTI-360 sequences used in Table 1. Then re-run the full pipeline on the first sequence with Metric3DV2 depths normalized by s_t before the G-ICP step, keeping the photometric refinement in Eq. (6) unchanged, and compare final PSNR/SSIM/LPIPS and Table 3-style pose errors. A second run can replace the depth maps with LiDAR depth for lifting and G-ICP. If both runs stay within about 0.2 dB and the pose errors remain near 0.001, the depth-scale concern is not load-bearing; if PSNR drops by more than 1 dB or pose errors grow by more than 10x, the central claim is contingent on Metric3DV2 and the reported pose table needs a definition of its alignment/normalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result is that a sequential, SfM-free pipeline keeps pose errors near zero and achieves >9 dB PSNR over CF-3DGS on 100-m KITTI-360 sequences. The linchpin is Eq. (3): relative poses are multiplied along the sequence, so any bias in pairwise transforms accumulates. Each pairwise transform is initialized by G-ICP on point clouds lifted from Metric3DV2 depth maps. G-ICP solves for a rigid transform between two point clouds assumed to be in the same metric scale; if Metric3DV2's per-frame scale differs by even 1-2%, the recovered translations are biased by that ratio and the photometric refinement in Eq. (6) must compensate against single-frame overfitted Gaussians. The paper gives no depth-error or scale-drift analysis on KITTI-360, despite LiDAR ground truth being available, and the ablation in Table 5 does not include a ground-truth-depth condition. Independently, the pose errors in Table 3 (average ATE of roughly 0.001) are implausibly small for a monocular sequential estimator over >100 m, suggesting an unreported normalization or Sim(3) alignment that makes the pose claim unfalsifiable as written. If depth-scale drift is non-negligible, both the pose numbers and the >9 dB rendering gain are partly inherited from the depth prior rather than demonstrated by the proposed ICP/voxel modules.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ICP-3DGS, an SfM-free pipeline for 3D Gaussian Splatting in large-scale unbounded scenes. For each adjacent frame pair, the method computes metric depth with the pretrained Metric3DV2 model, lifts the depth maps to point clouds, aligns them with G-ICP to obtain a coarse relative pose, and then refines that pose by optimizing an SE(3) transform on a per-frame overfitted Gaussian under an L1 photometric loss. A sky mask defined by the maximum depth value is used to exclude unreliable sky regions. The estimated relative poses are composed sequentially along the trajectory. Reconstruction uses a voxel-based densification strategy that adds points from new frames only in voxels where the current Gaussian point density is low relative to the new point cloud. Experiments on five selected KITTI-360 sequences and on Tanks and Temples report large PSNR gains over Nope-NeRF and CF-3DGS, low pose errors on KITTI-360, and an ablation showing contributions from the sky mask, G-ICP, and voxel densification.","tokens_in":8128,"tokens_out":4481,"duration_ms":49298,"significance":"If the reported results are correct, the paper makes a useful contribution: it is one of the first SfM-free 3DGS methods aimed at large outdoor scenes, it releases source code, and its ablation indicates that each proposed module contributes to the final rendering quality. The use of external benchmarks with fixed test frames and the comparison against two prior pose-free methods are strengths. However, the central claim depends on an unvalidated assumption about the metric-scale consistency of the pretrained depth model, and the evaluation lacks error bars, omits recent baseline methods, and reports pose errors that are implausibly small without a precise statement of the alignment protocol. These issues currently limit confidence in the magnitude of the reported gains and in the attribution of those gains to the proposed modules rather than to the depth prior.","major_comments":[{"comment":"The entire pose-estimation pipeline rests on the assumption that Metric3DV2 produces scale-consistent metric depth across frames. G-ICP between two depth-lifted point clouds recovers a rigid transform only if the clouds are in the same metric scale; any per-frame scale drift biases the recovered translation, and because Eq. (3) composes pairwise transforms, the bias accumulates along the trajectory. The paper provides no depth-error or scale-drift analysis on KITTI-360, where LiDAR ground truth is available, and Table 5 contains no ablation condition that replaces Metric3DV2 with ground-truth depth. Please add (i) per-frame depth MAE and scale error relative to LiDAR, (ii) pose and rendering results when the initialization uses ground-truth depth, and (iii) a controlled scale-perturbation experiment. Without these, the attribution of the 9 dB rendering gain and the low pose errors to the proposed ICP and voxel modules rather than to the depth prior is not established.","section":"§2.2, Eq. (3)"},{"comment":"The reported pose errors are implausibly small: an average ATE of roughly 0.001 m over trajectories exceeding 100 m is difficult to reconcile with a sequential monocular estimator. The text says that 'we align the scale and trajectory for Nope-NeRF and CF-3DGS before computing the camera pose error,' but it does not state whether the same alignment is applied to the proposed method or what exact alignment (e.g., Sim(3) Umeyama alignment) is used. Please specify the full evaluation protocol for all methods, report errors both before and after alignment, and include trajectory plots. As written, the pose claim is not falsifiable.","section":"§3.2, Table 3"},{"comment":"All quantitative results are single-run point estimates with no error bars, standard deviations, or statistical significance tests. The KITTI-360 evaluation uses only five sequences selected for mainly static traffic agents, which is a narrow subset of urban driving and does not exercise the method under moving vehicles, pedestrians, or changing lighting. Moreover, the introduction cites Free-SurGS and IncEventGS as relevant pose-free 3DGS methods, but neither is included in any comparison. Please report multiple runs with variance, evaluate on additional KITTI-360 sequences or the standard split, and compare against these recent baselines where applicable.","section":"§3.1, Tables 1-3"},{"comment":"Reproducibility-critical hyperparameters are missing. The voxel size and the density ratio threshold in §2.3 are not reported anywhere, the photometric loss weight λ in Eq. (5) is not specified, and the optimizer settings for the refinement in Eq. (6) are not described. A reader cannot reproduce Tables 1-3 or transfer the method to a new scene without guessing these values. Please report all hyperparameters and the training schedule.","section":"§2.3, §3.1"},{"comment":"The sky mask is defined as D_t = max(D_t), which assumes that the global maximum depth in every frame is always sky. In outdoor scenes, distant buildings, vegetation, or clipped depth values can also attain the maximum, and Metric3DV2's depth clamping may assign sky and other far regions the same maximum value. The paper does not quantify how well this mask agrees with true sky or compare it with a semantic sky segmenter. Please include a mask-quality evaluation and an ablation with a semantic sky mask to justify this design choice.","section":"§2.2, Eq. (7)"}],"minor_comments":[{"comment":"There are inconsistent spellings: 'SfM' in the title and abstract but 'SFM' in the body, 'out door' in the abstract, and 'Tanks and Temple' in Section 3.2. Please unify terminology and fix typos.","section":"Throughout"},{"comment":"The text says the Tanks and Temples qualitative results are 'presented in Figure 3', but the correct reference is Figure 4. The captions for Figures 3 and 4 are also nearly identical and should be made distinct.","section":"§3.3"},{"comment":"The sentence 'we achieve a significant 5.7 dB PSNR improvement on average' is ambiguous: 5.7 dB is the average improvement over Nope-NeRF on KITTI-360 (Table 1), not the improvement over CF-3DGS, while the surrounding text refers to both baselines. Please state the comparison explicitly.","section":"§3.2"},{"comment":"In Eq. (2), the view transform matrix W is used without first defining the coordinate frame convention. A brief definition of W and J would improve readability.","section":"§2.2, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is very short and reads like a conference paper rather than a full journal article, which may matter for fit with this journal's scope. One additional point for the editor: the depth model central to the method, Metric3DV2, is co-authored by one of the authors of this submission. This is not a circularity problem because Metric3DV2 is a publicly available, independently trained model, but I would recommend that the authors disclose the connection explicitly. The main technical risk is the unvalidated assumption of metric-scale consistency of the depth prior; the proposed evaluation additions (LiDAR-based depth error analysis and a ground-truth-depth ablation) are feasible and should be requested."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is a solid incremental engineering paper, not a field re-org. The method combines monocular depth (Metric3DV2) with G-ICP for relative pose initialization, adds a max-depth sky mask, and uses a voxel-density heuristic to densify Gaussians in under-covered regions. On the tested KITTI-360 subset it beats CF-3DGS by ~9 dB PSNR and Nope-NeRF by ~5.7 dB, and the ablations show each module earns its place. The code is promised. That is real value for people doing SfM-free outdoor reconstruction.\n\nThe soft spots are all around evidence. First, the pose errors in Table 3 are implausibly tight: ATE around 0.001 on sequences with 100+ m of displacement. For a sequential estimator built on monocular depth, that needs a direct explanation — whether Sim(3) alignment was applied to the method's own trajectory or only to baselines, and what the depth-scale drift of Metric3DV2 is on these sequences. The paper gives neither. Since G-ICP and the voxel densification both consume the depth maps, a missing ground-truth-depth ablation is a real gap. Second, the evaluation is narrow: five KITTI-360 sequences chosen to be mostly static, no error bars, no Free-SurGS or IncEventGS comparison, and key hyperparameters (voxel size, density ratio threshold, loss weight) are unreported. These are fixable, but as written they cap confidence.\n\nThe central claim is probably true inside the tested scope; the method does help. But the 'more than 9 dB' headline is selective (it is vs CF-3DGS only) and the pose numbers invite skepticism rather than closing it.\n\nWho is this for? People working on pose-free 3DGS and large-scale reconstruction. It deserves a serious referee, but that referee should send it back for a depth-scale analysis, pose-alignment details, hyperparameters, and a broader baseline set.\n\nRecommendation: accept for peer review, expect major revision.","headline":"Useful incremental SfM-free 3DGS pipeline with real gains, but the near-perfect pose errors and missing depth-scale analysis make the headline claim overreach as written.","tokens_in":8695,"tokens_out":3536,"would_cite":true,"duration_ms":37834,"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":"This paper proposes an SfM-free 3D Gaussian Splatting pipeline that uses monocular depth and Generalized ICP to recover camera poses and reconstruct large unbounded outdoor scenes without structure-from-motion.","keywords":["3D Gaussian Splatting","SfM-free","pose estimation","Iterative Closest Point","monocular depth","novel view synthesis","unbounded scenes","voxel densification"],"falsifier":"A decisive check would be to split KITTI-360 frames into high- and low-depth-error groups using the dataset's LiDAR ground truth, then measure the method's per-group ATE and PSNR; if the high-error group does not show substantially worse pose and rendering quality, the claimed dependence on monocular depth quality is cast into doubt, and if it does, the central assumption is confirmed and the method's range of applicability is precisely bounded.","tokens_in":7613,"feed_emoji":"🚗","tokens_out":8067,"duration_ms":70123,"temperature":0.7,"pith_summary":"The paper proposes a structure-from-motion-free 3D Gaussian Splatting pipeline that estimates camera poses and reconstructs a 3D scene from raw video, aimed at large-scale unbounded outdoor scenes where SfM preprocessing is brittle and expensive. The method uses monocular metric depth to lift each frame into a point cloud, aligns consecutive clouds with Generalized ICP to initialize the relative camera pose, and then refines that pose by optimizing Gaussians against the next frame; a sky mask removes unreliable infinite-depth regions, and a voxel-based densification step grows the Gaussian scene only where the current coverage is sparse. On KITTI-360 street sequences the authors report 23.32 dB PSNR against 13.96 dB for CF-3DGS and 17.60 dB for Nope-NeRF, with far lower trajectory error, and a smaller but consistent improvement on the bounded Tanks and Temples scenes. If those results stand, neural rendering from raw outdoor video becomes feasible without SfM, which matters because SfM is the conventional bottleneck in large-scale scene reconstruction.","feed_headline":"Pose-free Gaussian splatting jumps 9 dB on street scenes","feed_subtitle":"Monocular depth plus ICP estimates camera poses, so raw outdoor video renders without structure-from-motion.","key_machinery":"The load-bearing machinery is a camera-pose estimator built from monocular depth plus Generalized ICP, and a voxel-based scene-growth rule. For each adjacent frame pair, Metric3DV2 depth is lifted into two point clouds; G-ICP finds a coarse rigid transform between them, which seeds an optimization of a learnable SE(3) warp applied to the previous frame's Gaussians, supervised by an L1 rendering loss against the next frame. A sky mask defined by maximum depth excludes the unreliable infinite-depth background from this optimization. On the reconstruction side, the current frame's points are transformed into the scene coordinate frame and voxelized alongside the existing Gaussian centers, and new points are inserted only into voxels where the incoming density ratio exceeds a threshold, giving the scene a way to grow into areas that the first frame never covered.","core_discovery":"The central claim is that a 3D Gaussian Splatting scene can be built and its camera poses recovered from an unposed video sequence, even with large inter-frame motion and unbounded spatial extent, by pairing monocular metric depth with Generalized ICP for pose initialization and a voxel-driven growth rule for the Gaussian representation. In the proposed ICP-3DGS, each adjacent frame pair is registered by lifting both depth maps to point clouds and applying G-ICP to get a coarse relative transform; the transform then initializes a learnable SE(3) warp of the previous frame's Gaussians, refined with an L1 photometric loss against the next frame. Skipping sky pixels and comparing voxel occupancy between the new frame's points and the existing Gaussian centers lets the reconstruction expand into uncovered regions rather than relying on gradient-based splitting that stalls in empty space. The authors present this as an SfM-free 3DGS formulation that scales to street-level scenes, evidenced by 9 dB PSNR improvements over CF-3DGS on KITTI-360 and accurate recovered trajectories (average ATE of 0.001 after scale alignment).","pith_inferences":["We infer that a direct extension would be to feed the recovered poses and per-frame depth back into Metric3DV2 or a depth-refinement network, closing the loop between pose and depth; the ablations show a strong dependence on depth quality, so a depth-consistency loss across frames might push the pose error even lower.","We infer that the voxel density-ratio criterion could be made adaptive to local scene complexity, for example using finer voxels near the camera and coarser ones far away, which would plausibly improve the unbounded-scene results further than the fixed voxel size used here.","We infer that if the ATE numbers are reproduced, the output trajectories could plausibly replace ground-truth poses for training other view-synthesis or pose-estimation models in outdoor settings, a use the paper does not discuss.","We infer that the method inherits Metric3DV2's failure modes: on scenes with reflective or translucent surfaces, where monocular depth is notoriously unreliable, the pose initialization would be expected to degrade, a regime the paper's evaluations do not cover."],"forward_implications":["Adopting the pipeline would let practitioners reconstruct street-scale scenes directly from raw vehicle-mounted video, bypassing the feature-matching failure modes of SfM on low-texture or fast-moving footage.","The recovered camera trajectory is reported at least an order of magnitude more accurate than prior pose-free baselines (average ATE of 0.001 on KITTI-360 after alignment), which could make the pose output usable as a calibration or pseudo-ground-truth source in its own right.","The voxel-based densification means the method can grow into previously unseen scene regions, removing the assumption that the first frame's point cloud covers the whole reconstruction volume.","Because pose estimation and scene growth take roughly 1.5 hours on KITTI-360, comparable to CF-3DGS and far less than Nope-NeRF's 60 hours, the approach could make pose-free 3DGS the default for large-scale outdoor capture rather than a research novelty."],"supporting_citations":[{"why":"CF-3DGS: the direct baseline and the source of the Gaussian-level pose refinement formulation this work extends.","marker":"[8]"},{"why":"Metric3DV2: the pretrained monocular depth model whose metric depth maps supply the point clouds used for G-ICP and densification.","marker":"[13]"},{"why":"Generalized-ICP: the point-cloud registration algorithm that provides robust relative-pose initialization under large camera movements.","marker":"[12]"},{"why":"3D Gaussian Splatting: the explicit Gaussian scene representation and differentiable renderer the whole pipeline is built around.","marker":"[2]"},{"why":"Nope-NeRF: a leading pose-free NeRF baseline and the source of the train/test frame selection protocol used on both datasets.","marker":"[7]"},{"why":"Revisiting densification in Gaussian splatting: the analysis of gradient-based densification's failure in empty regions that motivates the voxel-based growth rule.","marker":"[16]"},{"why":"KITTI-360: the large-scale outdoor street dataset used for the main quantitative comparisons and trajectory evaluation.","marker":"[17]"}],"fun_headline_variants":["ICP-3DGS: SfM-free Gaussian splatting for large unbounded scenes","No SfM needed: ICP-3DGS does poses and voxel growth for 3DGS","9 dB better: Pose-free Gaussian splatting with ICP and voxels","Unposed video to 3D: ICP-3DGS scales to street scenes","ICP-3DGS: Replace SfM with ICP for large-scale 3DGS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes the pretrained monocular depth model Metric3DV2 produces sufficiently accurate, scale-consistent metric depth at every frame, because the depth lifts both the pose-initialization point clouds and the densification input; if that depth estimate drifts or fails, the pose initialization and scene growth collapse, and the reported gains evaporate.","fun_headline_variants_meta":{"raw":{"variants":["ICP-3DGS: SfM-free Gaussian splatting for large unbounded scenes","No SfM needed: ICP-3DGS does poses and voxel growth for 3DGS","9 dB better: Pose-free Gaussian splatting with ICP and voxels","Unposed video to 3D: ICP-3DGS scales to street scenes","ICP-3DGS: Replace SfM with ICP for large-scale 3DGS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1476,"prompt_tokens":935,"completion_tokens":541,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":423}},"tokens_in":551,"tokens_out":541,"duration_ms":5245,"temperature":1.0,"reasoning_tokens":423,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:58:21.788300+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check would be to split KITTI-360 frames into high- and low-depth-error groups using the dataset's LiDAR ground truth, then measure the method's per-group ATE and PSNR; if the high-error group does not show substantially worse pose and rendering quality, the claimed dependence on monocular depth quality is cast into doubt, and if it does, the central assumption is confirmed and the method's range of applicability is precisely bounded.","supporting_citations":[{"cited_title":"S mask” denotes the sky mask, and “VB-Den","cited_arxiv_id":null,"evidence_quote":"CF-3DGS: the direct baseline and the source of the Gaussian-level pose refinement formulation this work extends."},{"cited_title":"Nope-nerf: Optimising neural radiance field with no pose prior,","cited_arxiv_id":null,"evidence_quote":"Metric3DV2: the pretrained monocular depth model whose metric depth maps supply the point clouds used for G-ICP and densification."},{"cited_title":"Preliminary Problem formulation.Given a sequence ofNunposed RGB images frames, we denote it as{Ii}N i=1, whereI i ∈R H×W×3 and K is the camera intrinsics","cited_arxiv_id":null,"evidence_quote":"3D Gaussian Splatting: the explicit Gaussian scene representation and differentiable renderer the whole pipeline is built around."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Nope-NeRF: a leading pose-free NeRF baseline and the source of the train/test frame selection protocol used on both datasets."}],"review_version":1}