{"id":"30f6eea3-f19c-495e-a525-487cde9fde17","arxiv_id":"2607.23758","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A feed-forward Gaussian head on OmniVGGT plus road-plane grid fusion and structure-aware grouping reconstructs compact road surfaces that beat RoGS and AnySplat on Waymo and zero-shot nuScenes.","lead":"RoadVGGT turns multi-view driving images plus poses and depth into a compact Gaussian road map in one forward pass, without per-scene training. It matters because HD mapping and simulation need fast updates over newly collected roads rather than slow scene-by-scene optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline gains over AnySplat may be driven by input asymmetry, not method: RoadVGGT consumes GT poses, LiDAR depth, and external semantic masks, while the pose-free AnySplat baseline apparently receives none of these.","rationale":"The reader identified backbone-geometry dependence as the weakest assumption — a real limitation the authors themselves concede. I agree that is a concern, but I see a sharper, more load-bearing one for the strongest claim: the comparative tables that carry the \"improves over AnySplat/RoGS\" message appear to be run under asymmetric inputs. RoadVGGT gets poses, metric depth, and external semantics; the feed-forward baseline it is benchmarked against is designed for unconstrained, unposed views, and its 4.8 m Z-RMSE is consistent with missing metric scale rather than poor reconstruction. The mIoU comparison is similarly confounded because RoadVGGT's semantics are inherited, not predicted. This is not an accusation of impropriety — the paper is transparent that poses/depth/semantics are inputs, and the systems contribution (grid fusion, structure-aware grouping, compactness) stands on its own via the ablations in Table 2, which are internally consistent and well-designed. But the headline quantitative superiority claim needs a matched-input control to be trusted. The reader's CONDITIONAL verdict is the right level; I would keep it but redirect the condition from generic \"release code and stress-test noisy inputs\" to the specific matched-input rerun described above. Novelty and soundness assessments are unaffected.","tokens_in":11930,"tokens_out":2034,"duration_ms":57373,"concrete_test":"Run a matched-input comparison on the Waymo test split: (a) give AnySplat the same camera poses and LiDAR depth RoadVGGT receives (or align its output with a similarity transform and restrict metrics to road pixels via the same Mask2Former masks), and (b) ablate RoadVGGT by replacing provided depth/poses with backbone-predicted geometry only. If (a) closes most of the Z-RMSE and PSNR gap, or (b) drops RoadVGGT to AnySplat-level numbers, the headline improvement is input-conditioning rather than method; also recompute mIoU with all methods inheriting identical Mask2Former labels to isolate fusion quality from segmentation-source quality.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that RoadVGGT \"improves image quality, semantic mapping, and elevation accuracy relative to AnySplat and RoGS.\" But §3.1 defines the input as X = {I, K, T, D, S} — RGB, intrinsics, poses, depth observations, and Mask2Former semantic maps are all system inputs. AnySplat is explicitly a pose-free method \"from unconstrained views,\" and the paper never states that AnySplat was given poses, metric depth, or semantic masks. Two numbers in Table 1 make this smell load-bearing: (1) AnySplat's Waymo Z-RMSE is 4.8173 m versus 0.23/0.26 for the other two — a ~20x gap that is the signature of missing metric scale, not of worse reconstruction; if AnySplat's output is up-to-scale, the elevation comparison is nearly meaningless as run. (2) RoadVGGT's mIoU is majority-vote propagation of the external Mask2Former labels (§3.3: labels are \"auxiliary metadata,\" inherited per-pixel and voted per cell). Its 0.4692 mIoU therefore largely measures how faithfully the pipeline copies an external segmentation model into BEV; unless AnySplat and RoGS were fed the identical Mask2Former maps, the semantic-mapping win is confounded by input, not architecture. Even the PSNR gap (25.71 vs 24.09) is partly attributable to semantic-mask road filtering and metric-depth anchoring that the baseline lacked. If the improvements over the feed-forward baseline collapse under matched inputs, the paper's contribution narrows from \"better feed-forward road reconstruction\" to \"conditioning on GT geometry and semantics helps\" — a much weaker claim. The RoGS comparison is fairer in spirit (both use calibration), but there the margin is modest (25.71 vs 24.81 PSNR) and RoGS is handicapped by a preset lateral range, which the authors themselves note reduces its coverage.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper proposes RoadVGGT, a feed-forward pipeline for large-scale road surface reconstruction. It attaches a DPT-style Gaussian head to a frozen OmniVGGT backbone, which consumes multi-view RGB together with provided camera poses, LiDAR depth observations, and Mask2Former semantic maps. Predicted pixel-aligned 2D Gaussians are aligned into a metric world frame via Umeyama alignment, filtered to road-surface classes, and compressed by confidence-weighted fusion on a 0.05 m road-plane XY grid, with category-aware grouping and road–sidewalk junction protection to avoid destructive averaging of thin markings and adjacent surfaces. On Waymo (trained) and nuScenes (zero-shot), RoadVGGT reports better PSNR/SSIM/LPIPS, BEV mIoU, and Z-RMSE than AnySplat (feed-forward) and RoGS (per-scene optimization), at 112.81 MB storage and 0.84 s rendering (Table 1), with ablations for grid resolution, Gaussian parameterization, and grouping (Table 2). The central claims are (i) a feed-forward method can replace per-scene optimization for road mapping, and (ii) road-structure-aware fusion preserves quality while reducing storage.","tokens_in":12326,"tokens_out":3557,"duration_ms":78771,"significance":"If the comparison protocol is cleaned up, this is a useful and timely contribution: it is among the first works to convert geometric-foundation-model features into a compact, renderable road-surface map without per-scene optimization, and it shows zero-shot transfer from Waymo to nuScenes. Concrete strengths worth crediting: consistent gains over the per-scene-optimized RoGS (which plausibly received comparable privileged inputs) on all five quality metrics plus storage and runtime; a principled, physically meaningful metric XY-grid fusion whose cell size is comparable across trajectories; structure-aware grouping with a measurable elevation benefit; and a reasonably complete ablation suite (grid resolution at four scales, 2D vs 3D Gaussian parameterization, grouping components). The method is practically relevant for scalable HD-map updates. The significance is currently capped by the unverified fairness of the AnySplat comparison, which the abstract leans on.","major_comments":[{"comment":"§4.1/Table 1: the paper never states what inputs the baselines received. RoadVGGT's input is X={I,K,T,D,S} (§3.1): GT poses, LiDAR depth, and Mask2Former maps. AnySplat is a pose-free method 'from unconstrained views'; if it received no poses/depth, the comparison is confounded. Its Waymo Z-RMSE of 4.8173 m vs 0.23/0.26 for the posed methods is the signature of missing metric scale (median Z-offset alignment, §4.1, removes offset but not scale) — yet on nuScenes AnySplat reaches 0.5595, which is inconsistent with a purely up-to-scale output and is never explained. The manuscript must (a) state exactly what each baseline consumed, (b) describe how AnySplat outputs were scaled/aligned before Z-RMSE, and (c) ideally include a matched-input control (e.g., RoadVGGT without depth input, or AnySplat with GT-pose alignment) so the gain attributable to architecture vs. input conditioning is quant","section":"§4.1, Table 1"},{"comment":"RoadVGGT's semantic labels are not predicted: they are Mask2Former labels inherited per-pixel and majority-voted per grid cell (§3.1, §3.3). The Waymo mIoU gain (0.4692 vs 0.4361/0.2768) therefore largely measures faithful propagation of an external segmentation model. Two things must be clarified: (1) the provenance of the mIoU ground truth — if the reference BEV semantics derive from the same fine-tuned Mask2Former, RoadVGGT's score partially measures self-agreement with the label source; (2) whether AnySplat and RoGS were given the identical semantic maps. If the baselines used different (or no) label sources, the semantic-mapping column of Table 1 compares input pipelines, not reconstruction methods, and this caveat is load-bearing for the abstract's claim of improved 'semantic mapping'.","section":"§3.3, §4.1 (mIoU protocol)"},{"comment":"§4.1 states metrics are computed on the intersection of the reconstruction masks of the compared methods. This is fair per-pixel but hides coverage differences that the paper itself acknowledges (§4.3: RoadVGGT retains more visible surface than RoGS's preset lateral range). Two consequences: (i) a method reconstructing only easy central regions could score better on the intersection while being worse as a map; (ii) the storage comparison (112.81 MB vs AnySplat 712.47 / RoGS 178.67) compares representations of different spatial extents, since RoadVGGT's semantic filtering discards everything off-road while AnySplat models the full scene. Please report per-method reconstructed area/coverage, and either normalize storage by covered area or add a matched-coverage comparison. Per-scene variance (error bars or per-scene tables in the supplement) would also strengthen Table 1.","section":"§4.1 (evaluation protocol), Table 1 (Storage)"}],"minor_comments":[{"comment":"Table 2(a): the r=0.05 m default is described as the best balance, but it is not dominant on all metrics — r=0.02 m gives better mIoU (0.4850 vs 0.4692) and r=0.10 m gives better Z-RMSE (0.2040 vs 0.2312). The choice is defensible, but the text should acknowledge the trade-off explicitly rather than implying 0.05 m is uniformly best.","section":"Table 2(a)"},{"comment":"Since GT camera poses are provided to OmniVGGT, it is unclear why a full similarity (Umeyama) alignment of predicted vs. input trajectories is needed — the backbone already conditions on the input poses. Please clarify what residual misalignment this step corrects, and whether scale s deviates from 1 in practice.","section":"§3.2 (Umeyama alignment)"},{"comment":"Chunks are non-overlapping and processed independently, with a single global fusion afterward. Are seams or discontinuities visible at chunk boundaries in BEV or elevation maps? A brief comment or figure would help, since boundary artifacts are a natural failure mode of chunked inference.","section":"§3.5 (chunked inference)"},{"comment":"Figure 1 contains the typo 'Camera Paramters'; the author/affiliation block has a stray comma line (', cszhl@zju.edu.cn'); references are inconsistently formatted (mix of venue and arXiv citations, 'RoGs' vs 'RoGS'). Please also state whether code, the fine-tuned Mask2Former, and evaluation scripts will be released.","section":"Fig. 1, author block, References"},{"comment":"The junction-protection ablation (Table 2(c)) shows its main effect on Z-RMSE (0.2312 vs 0.2638) with negligible PSNR change; stating this explicitly would help readers understand which component buys which improvement. Similarly, the LPIPS weight λ=0.05 appears only in §4.1; defining it in §3.5 where the loss is introduced would improve readability.","section":"Table 2(c), §3.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid engineering contribution with unusually honest limitation disclosure (Conclusion). The main risk the editor should weigh is that the headline Table 1 improvements over AnySplat may be substantially input-driven rather than method-driven; the authors should be required to disclose exact baseline inputs and, ideally, add a matched-input comparison. There is no code-release statement; given the multi-author industry affiliation, the editor may wish to ask whether the fine-tuned Mask2Former, evaluation scripts, and scene lists can be released to support reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a clean systems paper that turns OmniVGGT into compact road-surface Gaussians without per-scene training, and the road-plane fusion with category/junction protection is the real design move. The AnySplat numbers are less clean than the abstract suggests.\n\nWhat is new is not the backbone or a Gaussian head—those are extensions of VGGT/OmniVGGT and feed-forward splat work—but the metric XY grid fusion plus structure-aware grouping aimed at long, flat road manifolds. They freeze the geometry stack, train only the head with L1+LPIPS after fusion, chunk long sequences, and get one compact representation that does RGB/semantic BEV, elevation, and NVS. Table 1 and the ablations (grid size, 2D vs 3D, grouping) are legible. Zero-shot nuScenes is a real plus. Storage drops hard versus dense prediction while quality holds at 0.05 m. That is useful for HD mapping people who hate per-scene RoGS-style training.\n\nSoft spots, in proportion. The stress-test on AnySplat mostly lands. RoadVGGT is fed poses, depth, and Mask2Former masks; AnySplat is a pose-free unconstrained-view method. The 4.8 m Z-RMSE on AnySplat looks like a scale mismatch, not a fair elevation contest, and mIoU is largely majority-vote copy of external labels, so the semantic win is partly input, not architecture. Against RoGS the comparison is fairer (both calibrated); margins are modest (PSNR 25.71 vs 24.81) but real, and coverage is better because they filter by semantics instead of a fixed lateral band. Authors themselves say quality still rides on the frozen backbone. No code, no error bars, free knobs (r, L_fine, chunk length). None of that sinks the paper; it narrows the claim from “beats feed-forward baselines” to “road-conditioned fusion on strong geometry priors works and removes test-time opt.”\n\nFor anyone building scalable road maps or simulation assets, this is worth a read. Math is standard, citations are in the right neighborhood, circularity is low. I would send it to referees; they should demand matched-input baselines and code. I would cite it if I were writing on feed-forward driving reconstruction. Bring it to reading group if the group cares about AD mapping; otherwise maybe.","headline":"Solid systems paper for feed-forward road Gaussians; the AnySplat comparison is partly confounded by inputs, but the RoGS win and road-plane fusion still make it worth engaging.","tokens_in":13538,"tokens_out":600,"would_cite":true,"duration_ms":10783,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Road surfaces can be rebuilt as compact Gaussians in one feed-forward pass, without per-scene training.","keywords":["road surface reconstruction","feed-forward Gaussian splatting","geometric foundation models","bird's-eye-view mapping","grid fusion","autonomous driving","elevation estimation"],"falsifier":"On held-out long road sequences, if feed-forward RoadVGGT (no test-time training) fails to match or beat a per-scene road Gaussian method on shared-region PSNR, BEV mIoU, and elevation error while staying more compact—or collapses when backbone depth/pose noise rises—the central claim does not hold.","tokens_in":13140,"feed_emoji":"🛣️","tokens_out":840,"duration_ms":19113,"temperature":0.7,"pith_summary":"Road mapping usually means optimizing a new model for every stretch of road, with coverage and hyperparameters tied to that drive. This paper argues that is unnecessary: a geometric foundation model, given multi-view images plus poses and depth, can drive a learned Gaussian head that predicts dense road attributes, which are then aligned into metric world coordinates and fused on a road-plane grid. Structure-aware grouping keeps thin markings and road–sidewalk boundaries from being averaged away. The result is one compact Gaussian road surface that yields RGB and semantic bird’s-eye maps, elevation, and novel views—without test-time scene optimization—and that beats both a generic feed-forward Gaussian baseline and a strong per-scene road optimizer on quality and storage, including zero-shot on a second driving dataset.","feed_headline":"Road maps from one pass—no per-scene training","feed_subtitle":"Metric plane fusion turns foundation-model Gaussians into compact BEV, elevation, and views.","key_machinery":"Road-structure-aware XY grid fusion: confidence-weighted merging of dense Gaussians on a metric road-aligned plane, with category-aware groups and road–sidewalk junction protection so only structure-compatible Gaussians fuse inside each cell.","core_discovery":"RoadVGGT shows that dense pixel-aligned Gaussian predictions from a geometric foundation model can be turned into a compact, renderable road surface by metric road-plane grid fusion with category-aware and road–sidewalk junction grouping, eliminating per-scene optimization while improving image quality, semantic mapping, and elevation accuracy relative to prior feed-forward and optimization-based road methods.","pith_inferences":["If backbone geometry is the bottleneck, joint or lightly adapted geometry training on road manifolds may unlock larger quality jumps than fusion tweaks alone.","The same plane-grid plus structure-group idea could extend to other near-manifold outdoor surfaces (parking lots, runways) where vertical extent is limited.","Chunk-then-global-fuse inference suggests a natural streaming mapper for continuous fleet collection, if drift across chunks is controlled."],"forward_implications":["New road logs can be mapped by direct inference instead of training a fresh scene model each time.","HD-map style products—RGB/semantic BEV, elevation, and novel views—can share one compact static Gaussian surface.","Gaussian capacity can scale with physical road extent via a metric plane grid rather than a fixed token budget or generic 3D voxels.","Cross-dataset zero-shot use becomes a practical target when poses and depth are available at capture time.","Gains in driving-scene geometric foundation models transfer immediately into better feed-forward road maps."],"fun_headline_variants":["Feed-forward Gaussians rebuild roads without per-scene training","RoadVGGT fuses foundation Gaussians into compact metric surfaces","One-pass road maps via road-aware Gaussian grid fusion","Compact BEV roads from pose-depth Gaussians—no test-time opt","Category-aware plane fusion yields scalable Gaussian road surfaces"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The frozen geometric backbone, given poses and depth, must already be accurate enough that a learned Gaussian head and fixed-grid fusion can recover thin markings and elevation without any per-scene refinement.","fun_headline_variants_meta":{"raw":{"variants":["Feed-forward Gaussians rebuild roads without per-scene training","RoadVGGT fuses foundation Gaussians into compact metric surfaces","One-pass road maps via road-aware Gaussian grid fusion","Compact BEV roads from pose-depth Gaussians—no test-time opt","Category-aware plane fusion yields scalable Gaussian road surfaces"]},"model":"grok-4.5","effort":"low","cost_usd":0.004116,"raw_usage":{"total_tokens":1255,"prompt_tokens":799,"num_sources_used":0,"completion_tokens":74,"cost_in_usd_ticks":41164000,"prompt_tokens_details":{"text_tokens":799,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":382,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":799,"tokens_out":74,"duration_ms":6899,"temperature":1.0,"reasoning_tokens":382,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T13:24:44.029001+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On held-out long road sequences, if feed-forward RoadVGGT (no test-time training) fails to match or beat a per-scene road Gaussian method on shared-region PSNR, BEV mIoU, and elevation error while staying more compact—or collapses when backbone depth/pose noise rises—the central claim does not hold.","supporting_citations":[],"review_version":1}