{"id":"dae3f29f-d7a1-4e92-a19e-0664df49d46b","arxiv_id":"2412.17612","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"CoSurfGS is a distributed device-edge-cloud training framework for 3D Gaussian surface reconstruction that compresses local models and distills them into a global large-scene model, reducing memory and training time.","lead":"CoSurfGS trains many small 3D Gaussian models on separate devices, then compresses and merges them into one global model of a large scene. The result is a pipeline that uses less GPU memory and less training time than a single centralized model, with surface geometry and rendering quality the paper says is state of the art.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LMC and MAS assume all local Gaussian models share one world coordinate frame, but Sec. 4.2 describes per-device SfM; no alignment or registration step is presented, so the fusion step may be merging models in incompatible coordinate systems.","rationale":"The reader's weakest assumption already pointed to the lack of explicit alignment or registration of overlapping surfaces at region boundaries. The stress-test pass identifies an even more basic version of the same problem: independent per-device SfM does not merely produce boundary inconsistency, it produces local reconstructions in different coordinate frames with arbitrary similarity transformations. Unless the implementation actually uses a common global pose set (which the text never states), the camera-overlap computation in Eq. 11 and the Gaussian merging in Sec. 4.4 are undefined. This directly threatens the central claim that LMC and MAS produce a high-quality global surface model. The paper contains no error analysis, no registration module, and no code release that would clarify the coordinate-frame situation. Because the published description is insufficient to establish that the distributed fusion is geometrically coherent, the verdict should remain conditional, with explicit requirement to document the coordinate-frame alignment or to evaluate the method under genuinely independent per-device SfM. The reader's conditional verdict is therefore confirmed, not overturned.","tokens_in":14199,"tokens_out":3811,"duration_ms":41115,"concrete_test":"Take the Rubble scene, split the images into two device partitions with substantial overlap, run COLMAP separately on each partition to obtain local camera poses and sparse points, and then feed these unaligned local models through the published LMC+MAS pipeline without any registration step. Measure PSNR and F-score against the reported numbers. If the local SfM coordinate frames differ by a nonzero similarity transform and the pipeline still matches the reported quality, the concern is refuted; if fusion fails or quality collapses, the method depends on an unstated global alignment (e.g., using the dataset's global COLMAP poses) and the paper must specify and evaluate that condition.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim depends on the fusion of independently trained local Gaussian models into one coherent global surface. The paper states in Sec. 4.2 that each device captures images and 'performing SfM to initialize both extrinsic and intrinsic' for its own area. If each device runs SfM independently, the resulting local reconstructions have arbitrary similarity gauge freedom (rotation, translation, scale) relative to each other. No global registration, loop closure, or common-world-frame constraint is described anywhere in the method. This breaks both key fusion components: LMC's camera-overlap ratio in Eq. 11 requires comparing local camera sets C_l and global camera sets C_g, which is geometrically meaningless if their extrinsics live in different coordinate frames; and MAS's initialization by directly merging compressed local Gaussians, followed by self-distillation in Eq. 13, assumes the merged points already occupy a common world space. If the local frames disagree, the same physical surface appears as duplicated or offset Gaussian clouds, and the distillation losses L_d and L_n (Eq. 14) will average inconsistent depth and normal maps rather than align geometry. The reported experiments on MegaNeRF and UrbanScene3D likely use precomputed global poses from the datasets, which would supply a common frame, but the paper does not state this, and its device-side SfM description suggests otherwise. This is a more fundamental issue than boundary-region geometric disagreement: without coordinate alignment, the entire distributed fusion procedure is not well-defined as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"CoSurfGS proposes a device-edge-cloud distributed learning framework for large-scale 3D surface reconstruction with Gaussian splatting. Each device trains a local PGSR-style flattened-Gaussian model, followed by Local Model Compression (LMC), which prunes Gaussians according to camera-overlap ratios and a priority score, and a Model Aggregation Scheme (MAS), which merges compressed local models and refines the global model via self-distillation from local RGB, depth, and normal renders. Experiments on BlendedMVS, Mill19, and UrbanScene3D report improved mesh F-scores, faster training times, and lower memory consumption relative to centralized and distributed baselines.","tokens_in":14517,"tokens_out":4494,"duration_ms":43157,"significance":"The work addresses a practically important problem: making 3DGS surface reconstruction feasible for large scenes under limited per-device memory. The distributed architecture is novel and the ablations in Tables 4 and 5 show that both LMC and MAS contribute to performance. If the coordinate-frame assumption is clarified and the experimental claims are appropriately qualified, the framework could be a useful contribution to large-scale reconstruction. The paper does not provide code, formal guarantees, or error analysis, so its significance rests mainly on the empirical validation, which currently has important gaps.","major_comments":[{"comment":"The paper states in the Figure 2 caption that each device performs its own SfM to initialize extrinsics and intrinsics. If each device runs SfM independently, the resulting local Gaussian models live in arbitrary similarity gauges (rotation, translation, and scale) that are not aligned to a common world frame. The LMC camera-overlap ratio in Eq. 11 and the MAS merging and self-distillation in Eq. 13 both presuppose a common coordinate frame; without an explicit alignment or registration step, the fusion is geometrically invalid. Even if the datasets provide globally consistent poses (as is typical for Mill19 and UrbanScene3D), the manuscript never states this and the device-side SfM description directly contradicts it. This is the central load-bearing issue and must be resolved, either by describing a global pose initialization or by adding an alignment step.","section":"Sec. 4.2, Fig. 2; Sec. 4.3 Eq. (11); Sec. 4.4 Eq. (13)"},{"comment":"The contributions claim that the method 'surpasses all existing methods' in surface reconstruction and the abstract claims strong photorealistic rendering. In Table 2, however, the no-mesh baseline DOGS achieves higher PSNR than CoSurfGS on 4 of 5 scenes (Building 22.73 vs 22.40, Rubble 25.78 vs 25.39, Campus 24.01 vs 23.63, Sci-Art 24.42 vs 23.29). The state-of-the-art claim needs to be restricted to the 'With Mesh' surface-reconstruction group, and the text should be revised so that the 'comparable to NVS methods' statement is accurate with respect to PSNR.","section":"Abstract, Contributions, and Table 2"},{"comment":"All reported metrics are from single runs with no error bars or variance estimation. Some differences are very small, e.g., the Scene-03 F-score gap of 0.0032 in Table 1. Additionally, the F-score error margins in the supplementary (0.5, 0.1, 0.5, 0.2 meters for the four BlendedMVS scenes) vary across scenes without justification, which hampers fair comparison. The quantitative support for the central claim would be substantially stronger with multiple seeds, standard deviations, or at least an explicit acknowledgment that the reported gaps may be within run-to-run noise.","section":"Tables 1, 2, 4, 5, 6; Supp. 8.2"},{"comment":"The paper does not analyze the case where overlapping local models disagree on surface geometry. The self-distillation loss in Eq. 13, with depth and normal supervision in Eq. 14, averages the teacher predictions; if the local surfaces are inconsistent near boundaries, the global model may average away fine structure instead of aligning it. The authors should provide a quantitative boundary-region analysis (e.g., geometry error as a function of distance to partition boundaries) or an experiment that isolates the effect of inconsistent local geometry. This is a broader correctness-risk concern that is not addressed by the current qualitative boundary visualizations.","section":"Sec. 4.3 and 4.4"}],"minor_comments":[{"comment":"Fed3DGS appears in the qualitative comparison in Fig. 4 but is not included in any quantitative table; either provide its results or remove it from the comparison.","section":"Fig. 4 and Tables"},{"comment":"The notation 'S_pro,k = M HWX' is unclear; the sum index and its range should be spelled out properly, and the definition of V(Σ) should be made less ambiguous.","section":"Eq. (10)"},{"comment":"The supplementary material contains numerous typos and formatting errors, including 'Visionlization', 'comaprison', 'mehtod', 'acheives', 'campare', 'noraml', and the corrupted percentage markers in Table 7. The manuscript should be carefully proofread.","section":"Supp. Sec. 9 and Tab. 7"},{"comment":"The qualitative text says 'the only method that can model the detail of the telegraph pole' but this claim is not supported by a quantitative metric; consider adding a focused detail-evaluation metric or tempering the claim.","section":"Sec. 5.2 and Table 2"},{"comment":"The privacy statement ('by only uploading Gaussian models instead of raw images') is plausible but not evaluated; if privacy preservation is a claimed contribution, the paper should at least discuss model inversion resilience or refer to previous attacks.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The coordinate-frame ambiguity is the most serious issue. If the authors confirm that global poses from the datasets are used and clarify the device-side SfM description, the main technical concern would be reframed. However, the state-of-the-art claim and the absence of error bars need to be addressed before acceptance. The paper's strong reliance on PGSR, co-authored by several of the same researchers, is worth watching: the comparison against PGSR+VastGS is reasonable, but a more independent baseline set (e.g., DOGS with mesh extraction, Fed3DGS) would strengthen the validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea is new and worth thinking about—distributed 3DGS surface reconstruction via device-edge-cloud training, with local compression and self-distillation aggregation. The ablations suggest LMC and MAS each do something. But the paper omits a load-bearing detail: how the independently trained local models end up in a common world frame. The text says each device runs its own SfM (Sec. 4.2). If that is literal, then the local Gaussians, poses, depth, and normal maps are in arbitrary similarity frames, and Eq. 11's camera overlap and Eq. 13's distillation are geometrically meaningless. The datasets (MegaNeRF, UrbanScene3D) ship global COLMAP poses, so the experiments may well use those; the paper just never says so. This is the first thing I'd ask the authors to fix.\n\nWhat's genuinely good: the framing as a distributed surface-reconstruction problem rather than just rendering is a real step beyond VastGaussian/DOGS/Fed3DGS. The LMC pruning criterion (priority score based on projection frequency, opacity, volume) is reasonable, and the MAS self-distillation as a boundary-blur fix is a sensible choice. The reported time/memory savings are large and the qualitative visuals look plausible.\n\nWeaknesses beyond the coordinate issue: no error bars, single runs; the F-score thresholds differ across BlendedMVS scenes (0.5, 0.1, 0.5, 0.2 m) without discussion; the claim of 'surpassing all existing methods' is too strong because DOGS beats them on PSNR/SSIM for some rendering metrics; no code or data release. The partition ablation is helpful but doesn't test boundary coherence directly.\n\nNet: this deserves a serious referee—the problem is timely and the approach is mostly coherent—but the current version needs a major revision that clarifies the global frame assumption and tightens the evaluation. I'd not cite it as-is, but I'd read a revised version carefully.","headline":"A plausible distributed surface-GS framework, but the paper never explains how independently trained local models share one coordinate frame—without that, the fusion step is not well-defined.","tokens_in":15090,"tokens_out":2167,"would_cite":false,"duration_ms":20252,"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 claims that distributed local Gaussian models, compressed and aggregated by self-distillation, reconstruct large 3D surfaces with less memory and time than centralized training.","keywords":["3D Gaussian splatting","surface reconstruction","distributed learning","large-scale reconstruction","model compression","self-distillation","device-edge-cloud architecture","multi-agent collaboration"],"falsifier":"Take a planar wall that spans two devices, each trained on its own half with a narrow shared strip of cameras. After LMC plus MAS fusion, extract the mesh and measure its deviation from a laser scan of the wall; if the strip shows a systematic kink, double surface, or inflated error while each half's interior is accurate, the assumption that self-distillation aligns rather than averages boundary geometry is broken.","tokens_in":30,"feed_emoji":"🏙️","tokens_out":8101,"duration_ms":119096,"temperature":0.7,"pith_summary":"CoSurfGS tries to make high-quality 3D surface reconstruction of large scenes practical by distributing the work across many devices instead of training one global model on a single GPU. It claims that a device-edge-cloud pipeline, in which each device trains a 3D Gaussian surface model on its own image subset, can be stitched into a single global model through two operations: local model compression (LMC), which prunes redundant boundary Gaussians, and a model aggregation scheme (MAS), which uses the local models as teachers to refine the global model. The paper reports that this reduces training time and GPU memory compared with centralized large-scene baselines while improving surface-reconstruction accuracy measured by mesh F-score, and stays competitive in novel-view rendering. A sympathetic reader would care because large outdoor scenes currently push surface-reconstruction methods past single-GPU memory limits, and the proposed pipeline converts that bottleneck into parallel, privacy-preserving local training.","feed_headline":"Distributed Gaussians cut memory and speed large-scene reconstruction","feed_subtitle":"Local models train in parallel, get pruned, then fuse by self-distillation into one global surface model.","key_machinery":"The load-bearing machinery is the two-stage local-to-global transition. LMC computes a per-Gaussian priority score from the frequency with which the Gaussian is hit by rays, its opacity, and a normalized volume factor, then prunes the fraction determined by how many cameras the local model shares with the global model, so only Gaussians with reliable multi-view agreement survive. MAS replaces naive set union with self-distillation: the compressed local models act as teachers, rendering RGB, depth, and normal maps from their own camera poses, and the global model is optimized against those teacher renderings plus its own flatness and single-view geometry losses, with no densification. Because the same LMC plus MAS pair is applied at the device-edge and edge-cloud levels, the scheme composes into a hierarchy that can scale the reconstructed region without changing the core operations.","core_discovery":"The paper's central claim is that a large scene's surface can be reconstructed more accurately and far more cheaply by converting the global optimization into a two-level distributed procedure. Each device trains a local planar-surface Gaussian model, following the PGSR approach, with flattening and single- and multi-view geometric constraints; before uploading, LMC prunes Gaussians that appear in few cameras, have low opacity, or sit in regions already covered by other models, using a priority score and a camera-overlap ratio. The edge or cloud then initializes a global model from the compressed local models and refines it with MAS, a self-distillation step where each local model renders RGB, depth, and normal images that supervise the global model, so the global surface is pulled toward the geometry each local model actually saw. The paper argues this avoids the blurry boundaries produced by naive unions of local models without expanding training regions or sharing neighboring images, and presents evaluations on aerial and MVS datasets showing better mesh F-scores, faster training, and lower memory than centralized and prior distributed methods.","pith_inferences":["Beyond the paper, the same teacher-student aggregation could be applied to implicit neural surface representations, where local signed-distance or occupancy fields play the role of the compressed Gaussian teachers, if a differentiable renderer supplies depth and normal supervision.","The priority score used by LMC is a proxy for multi-view consistency; a testable extension is to replace it with a direct photometric consistency measure, for example reprojection error against neighboring views, and measure whether boundary artifacts shrink further.","A natural extension is to measure geometric error in a strip around partition seams separately from interior error, since the paper only reports whole-scene F-scores; this would show whether MAS aligns or averages boundary geometry.","The ablations show more partition blocks slightly reduce quality but save time, suggesting an adaptive partition that aligns block boundaries with low-texture or weakly observed regions could recover the quality loss while keeping the speed."],"forward_implications":["Large scenes that previously caused out-of-memory errors on a single GPU can be reconstructed on consumer hardware by splitting the work among devices and aggregating only the compressed predictions.","Training time is cut substantially because all local models train in parallel and the global refinement lasts only a few distillation epochs, making iterative, practical use of large-scene reconstruction feasible.","Because only Gaussian models are uploaded and raw images stay on each device, the pipeline offers a privacy-preserving route to collaborative scene modeling across data owners.","The local-first strategy yields better surface geometry than a single global model, since each device's model converges on a smaller region and then passes its geometric knowledge to the global model through depth and normal supervision.","The same device-edge-cloud aggregation can be repeated over more hierarchy levels, so the method scales beyond the tested scene sizes by adding more devices or edges."],"supporting_citations":[{"why":"Supplies the 3D Gaussian primitive representation, densification, and reconstruction loss that each local model starts from.","marker":"[14]"},{"why":"Provides the planar Gaussian flattening and the single- and multi-view geometric losses used to keep local surfaces accurate.","marker":"[4]"},{"why":"Defines the distributed large-scene Gaussian baseline whose boundary-expansion and trimming strategy the paper's no-expansion MAS is built to replace.","marker":"[18]"},{"why":"Distributed Gaussian baseline with a Gaussian-consensus aggregation, used as a comparison for aggregation quality and resource use.","marker":"[5]"},{"why":"Supplies the Mill-19 large-scale dataset and the Mega-NeRF baseline used for novel-view-synthesis comparisons.","marker":"[29]"},{"why":"Surface-aligned Gaussian splatting baseline compared on mesh and rendering quality.","marker":"[8]"},{"why":"Neural surface reconstruction baseline compared on large-scene mesh accuracy.","marker":"[17]"},{"why":"Neural implicit surface baseline compared on surface reconstruction and rendering.","marker":"[32]"},{"why":"Supplies the BlendedMVS test scenes on which the surface reconstruction F-scores are measured.","marker":"[39]"}],"fun_headline_variants":["Distributed Gaussians scale surface reconstruction to large scenes","Collaborative splatting cuts GPU memory for big-scene 3D","Parallel Gaussian models fuse into one fast surface","Multi-agent Gaussians: large scene reconstruction with less memory","Local models, global surface: distributed 3D reconstruction"],"cache_read_input_tokens":17152,"weakest_assumption_plain":"The whole pipeline depends on local models trained on separate chunks of the scene agreeing well enough at their borders that pruning and self-distillation can stitch them into one coherent surface without any explicit geometric alignment.","fun_headline_variants_meta":{"raw":{"variants":["Distributed Gaussians scale surface reconstruction to large scenes","Collaborative splatting cuts GPU memory for big-scene 3D","Parallel Gaussian models fuse into one fast surface","Multi-agent Gaussians: large scene reconstruction with less memory","Local models, global surface: distributed 3D reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1271,"prompt_tokens":935,"completion_tokens":336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":255}},"tokens_in":551,"tokens_out":336,"duration_ms":3876,"temperature":1.0,"reasoning_tokens":255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:20:43.161820+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a planar wall that spans two devices, each trained on its own half with a narrow shared strip of cameras. After LMC plus MAS fusion, extract the mesh and measure its deviation from a laser scan of the wall; if the strip shows a systematic kink, double surface, or inflated error while each half's interior is accurate, the assumption that self-distillation aligns rather than averages boundary geometry is broken.","supporting_citations":[{"cited_title":"Mega-nerf: Scalable construction of large- scale nerfs for virtual fly-throughs","cited_arxiv_id":null,"evidence_quote":"Supplies the Mill-19 large-scale dataset and the Mega-NeRF baseline used for novel-view-synthesis comparisons."},{"cited_title":"Neuralangelo: High-fidelity neural surface reconstruction","cited_arxiv_id":null,"evidence_quote":"Neural surface reconstruction baseline compared on large-scene mesh accuracy."},{"cited_title":"Blendedmvs: A large- scale dataset for generalized multi-view stereo networks","cited_arxiv_id":null,"evidence_quote":"Supplies the BlendedMVS test scenes on which the surface reconstruction F-scores are measured."}],"review_version":1}