{"id":"ebf49fc7-267a-4b02-9656-d2b7171abdb4","arxiv_id":"2608.10682","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"VGGD improves single-frame surround-view driving reconstruction by feeding frozen VGGT geometry tokens into a dual-path neck and scale warmup before pixel-volume Gaussian decoding, reaching PSNR 24.85 on nuScenes.","lead":"This paper describes VGGD, a system that adds a frozen pretrained visual-geometry model to a Gaussian-splatting pipeline so that six surround-view car cameras can reconstruct a 3D driving scene in a single forward pass. It reports the best rendering scores on the nuScenes single-frame benchmark and slightly better depth correlation than prior work.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The geometry-consistency headline is not independently grounded: rendered depth is trained against Metric3D-v2 pseudo-depth (Eqs.","rationale":"The reader's weakest assumption already identifies the pseudo-depth supervision/evaluation loop, and my stress-test confirms that this is the most load-bearing concern. The rendering claim is plausible but small; the geometric-consistency claim is the part that must justify the 'geometry foundation-aware' framing, and it rests on a metric that is both trained toward and evaluated against monocular depth priors. PCC's invariance to scale specifically undermines the scale-drift narrative of Scale Warmup, making the concern concrete rather than merely a call for more data. A LiDAR- or triangulation-based validation would settle whether the PCC improvement reflects true 3D structure or learned alignment to monocular depth biases. Since the reader already conditioned acceptance on exactly this kind of validation, the verdict should remain CONDITIONAL.","tokens_in":15360,"tokens_out":5451,"duration_ms":59205,"concrete_test":"On a held-out nuScenes validation subset, accumulate LiDAR sweeps over a short time window, project them into the six target camera views at both endpoint ego poses, and compute for VGGD and Omni-Scene (a) affine-invariant RMSE and delta-1 accuracy after per-image scale/shift alignment and (b) cross-view consistency as the Chamfer distance between LiDAR points and back-projected rendered depth within the cuboid range B. If VGGD does not beat Omni-Scene on these LiDAR-based metrics, the claimed geometric-consistency and scale-warmup benefits collapse to alignment with monocular depth priors rather than true 3D consistency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim has two advertised improvements: novel-view rendering quality and geometric consistency. The rendering gain is modest (+0.58 PSNR over Omni-Scene) and lacks error bars, which is a reproducibility concern but not a conceptual flaw. The geometric-consistency claim is what carries the 'geometry foundation-aware' thesis, and it rests entirely on PCC evaluated against Depth Anything V2 pseudo relative depth (Sec. V-A.2). Meanwhile, the same geometry pathway is explicitly trained to match Metric3D-v2 coarse metric pseudo-depth: Eq. 7 applies a warmup L1 loss directly to the predicted geometry-path depth, and Eq. 9 includes rendered-depth losses L_dep against the same family of monocular depth estimates (Sec. V-A.4). That creates a circularity risk: the evaluation metric rewards rendered depth maps that resemble off-the-shelf monocular depth estimates, which the training objective already encourages. Because PCC is invariant to global shift and positive scaling, it also cannot detect the metric-scale drift that Scale Warmup is claimed to suppress; a globally wrong scale can yield near-perfect PCC. The qualitative depth comparisons in Figures 3-5 likewise contain no metric ground truth. Thus the geometric-consistency contribution is underdetermined by the evidence as presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes VGGD, a feed-forward 3D Gaussian Splatting method for single-frame surround-view driving reconstruction on nuScenes. VGGD takes six surround images, extracts frozen VGGT multi-view tokens, passes them through a dual-path neck that separates geometry and appearance features, predicts per-view depth with a transient Scale Warmup, and decodes a hybrid pixel-volume 3DGS scene. The paper reports state-of-the-art rendering quality (PSNR 24.85, SSIM 0.758, LPIPS 0.216) and improved relative geometric consistency (PCC 0.808) over Omni-Scene and other baselines, with ablations supporting Scale Warmup and the dual-path design.","tokens_in":15592,"tokens_out":4827,"duration_ms":47367,"significance":"The main contribution is architectural: shifting geometric reasoning from the decoder to a frozen VGGT frontend with driving-oriented adaptation. If the reported results are robust, this is a useful direction for sparse surround-view reconstruction, and the paper is transparent about its training objective, the pseudo-depth references, and the exact loss weights. The rendering comparisons use standard external benchmarks, and the internal ablations are coherent and consistent with the paper's narrative. However, the geometric-consistency headline is supported only by a pseudo-depth correlation, and the ablation evidence is obtained at a much shorter training budget than the final model. The significance is therefore conditional on stronger geometric validation and on demonstrating that the small gains over Omni-Scene are not within run-to-run variability.","major_comments":[{"comment":"The geometric-consistency claim is evaluated with PCC against Depth Anything V2 pseudo relative depth, while the same geometry pathway is trained against Metric3D-v2 pseudo-depth. This creates a circularity risk: the evaluation metric rewards rendered depths that resemble an off-the-shelf monocular depth prior, not metric or multi-view 3D consistency. PCC is invariant to global shift and positive scaling, so the claimed suppression of scale drift by Scale Warmup cannot be validated by PCC. I request validation against LiDAR point clouds or multi-view triangulated geometry, using scale-sensitive metrics such as absolute/relative depth error after alignment or reprojection error, or at minimum an explicit demonstration that the gains persist under scale-sensitive metrics.","section":"Sec. V-A.2, Sec. V-A.4, Eqs. (7) and (9)"},{"comment":"All ablations are trained for 10k iterations with a different final learning rate (1e-5) than the final model (100k iterations, 1e-6). Consequently Table II validates components only in the short-training regime; it does not show that the gains of Scale Warmup or Dual-Path Neck persist at the reported deployment budget. For example, SW1-None at 10k yields 21.92 PSNR, but nothing in the paper rules out that longer training closes the gap. The ablations should be rerun at the full 100k schedule, or the main table should also be reported at the ablation budget.","section":"Sec. V-D.1 and Tables I-II"},{"comment":"No error bars or repeated-seed statistics are reported, and the headline gains over Omni-Scene are small (PSNR +0.58, SSIM +0.022, LPIPS -0.021, PCC +0.004). Given that the PCC difference is four-thousandths on a pseudo-depth correlation, the claim of improved geometric consistency is not supported as statistically meaningful. Please report standard deviations over multiple training seeds and, for the PCC difference, a significance test or confidence interval.","section":"Table I"}],"minor_comments":[{"comment":"The dimensions and downsampling of the three VGGT token types are not given, and the adapter architecture for Ψgeo and Ψapp is described only in words; please specify these details for reproducibility.","section":"Sec. IV-B"},{"comment":"The qualitative panels labeled Geo1–Geo3 and App1–App3 are not mapped to any variant in Table II, so the figure cannot be interpreted unambiguously.","section":"Fig. 5 and Sec. V-D"},{"comment":"The coarse depth reference \\tilde D^v_t is introduced before its generation is described; the text only clarifies in Sec. V-A.4 that Metric3D-v2 pseudo-depth is used. Please move that definition earlier or add an explicit pointer.","section":"Algorithm 1 and Eq. (7)"},{"comment":"The footnote says STORM is included as a spatio-temporal reference; the authors should state whether STORM was retrained or evaluated under the same single-frame protocol as the other baselines, since STORM's native input is spatio-temporal.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope and the engineering is competent, but the novelty is incremental over Omni-Scene with VGGT features, and the pseudo-depth evaluation loop is the main risk. The editor should also weigh that the central idea is close to the authors' own VGD preprint, which is cited but not differentiated in detail. If the authors provide LiDAR-based or triangulation-based geometric validation and full-budget ablations, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"VGGD is a straightforward extension of the Omni-Scene/DrivingForward line: freeze VGGT, add a dual-path neck to keep appearance features from being crushed by geometry tokens, and add a transient depth warmup to stabilize early scale learning. The rendering numbers on nuScenes are best-in-table (24.85 PSNR vs 24.27 for Omni-Scene), and the ablations are internally consistent. I especially credit the warmup ablation: dropping it costs more than 2 dB PSNR in the 10k-iteration setup, which makes the scale-anchoring story concrete rather than hand-wavy.\n\nWhere the paper gets soft is the geometric-consistency claim. PCC is computed against Depth Anything V2 pseudo-depth, and the same geometry pathway is trained against Metric3D-v2 pseudo-depth. That is a real circularity risk: the model is rewarded for producing depth that resembles off-the-shelf monocular estimates. PCC being invariant to scale makes it even weaker for validating the scale-warmup benefit. The authors need a LiDAR-based or multi-view triangulated depth evaluation to back the 'improved geometric consistency' sentence.\n\nMinor but real: there are no error bars anywhere, so the +0.58 PSNR over Omni-Scene could be noise. The ablations run for 10k iterations while the final model runs 100k; that's common, but it means the warmup schedule's interaction with long training isn't fully controlled. And VGD, the closest prior from the same group, is cited but not included in the comparison table. If the authors are claiming SOTA among feed-forward surround-view methods, omitting the nearest baseline is a hole. The code link is a promise, not a release.\n\nBottom line: this is a coherent, useful empirical paper. The rendering improvement is modest but the architecture direction is plausible. It deserves a serious referee, but the review should require (1) code and checkpoints, (2) error bars or significance, and (3) geometry validation against something that isn't itself a monocular depth prior. Readership: people working on feed-forward 3D reconstruction for driving simulation.","headline":"A solid incremental surround-view reconstruction paper whose rendering gain is believable but whose geometric-consistency claim rests on pseudo-depth circularity.","tokens_in":16153,"tokens_out":3335,"would_cite":true,"duration_ms":30376,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"VGGD shifts geometric reasoning to a frozen VGGT frontend and reports the best rendering quality on the nuScenes single-frame surround-view benchmark.","keywords":["3D Gaussian Splatting","surround-view reconstruction","geometry foundation models","novel view synthesis","autonomous driving","nuScenes","feature decoupling","scale warmup"],"falsifier":"Evaluate rendered depth consistency against LiDAR point clouds or multi-view triangulated ground truth on nuScenes instead of pseudo relative depth. If VGGD's PCC advantage over Omni-Scene shrinks or reverses under true metric evaluation, or if removing Scale Warmup no longer hurts when trained with LiDAR labels, the paper's geometric-consistency claim is not supported.","tokens_in":15148,"feed_emoji":"🚗","tokens_out":4095,"duration_ms":37186,"temperature":0.7,"pith_summary":"The paper argues that in single-frame surround-view driving reconstruction, where six cameras barely overlap, the bottleneck is not the Gaussian decoder but the geometric quality of upstream features. It proposes VGGD, which feeds a frozen visual geometry foundation (VGGT) into a dual-path neck that separates geometry-consistent from appearance-aware features, applies a transient scale warmup to anchor early depth learning, and renders with a hybrid pixel–volume 3D Gaussian head. On the nuScenes single-frame benchmark, VGGD reports the best overall novel-view quality (PSNR 24.85, SSIM 0.758, LPIPS 0.216) and slightly higher relative depth consistency (PCC 0.808) than prior driving-centric and generalizable methods. The paper's claim is that geometry-aware upstream representations, adapted to the driving camera setting, relieve the ill-posedness of minimal-overlap 3D inference better than decoder-side corrections alone.","feed_headline":"Frozen geometry frontend tops single-frame surround-view rendering","feed_subtitle":"A pretrained VGGT backbone plus a two-path neck lifts PSNR and depth consistency over decoder-heavy rivals.","key_machinery":"The load-bearing objects are: the frozen VGGT geometry foundation, whose frame and global tokens encode cross-view structure and whose DINO-aligned tokens retain local appearance; the Dual-Path Neck, which routes geometry-token-only features to depth and geometry prediction while routing full tokens to appearance features, then fuses the two streams by addition; and Scale Warmup, a gated L1 loss on predicted depth active only for the first S_warm training steps, preventing scale drift under ego-pose changes without permanently constraining the depth network. These feed a hybrid pixel–volume 3DGS decoder: the pixel head consumes full points and features for image-aligned detail and far-range content, while the volume head consumes a masked, driving-centric bounded subset for 3D completion. The mechanism's role is to make the 3D support of the Gaussians reliable before any rendering loss is applied.","core_discovery":"The central claim is that sparse surround-view reconstruction should shift geometric modeling from the decoding head to a pretrained geometry foundation frontend. VGGT, frozen, supplies per-view DINO-aligned, frame, and global tokens that carry transferable multi-view structure; a Dual-Path Neck then decouples a geometry-consistent path, which predicts depth and 3D support, from an appearance-aware path, which preserves texture and semantic cues for weakly observed regions. A transient Scale Warmup supervises early depth with a coarse reference for the first 2,000 steps, anchoring metric scale before end-to-end rendering refinement takes over. The hybrid pixel–volume Gaussian decoder, taken from the Omni-Scene line but fed directly by both neck paths, converts the adapted features and back-projected points into a renderable 3D Gaussian scene. The paper reports that this combination beats per-scene optimization methods, general-purpose feed-forward models, and prior driving-centric pipelines on the nuScenes single-frame protocol, with the largest gains coming when warmup and both neck paths are present.","pith_inferences":["Because the method freezes VGGT, the same adapter could in principle be attached to newer geometry foundations; if their tokens are better calibrated to driving cameras, the warmup schedule may shorten or disappear.","The PCC evaluation is correlational and shift-invariant, so the reported geometric-consistency gains do not establish metric-scale accuracy; a LiDAR-based evaluation would test whether the warmup actually produces metric depth.","The dual-path design suggests a more general recipe for sparse-view reconstruction: keep a geometry-dedicated feature stream and a separate appearance stream, and fuse late, rather than forcing one representation to serve both roles.","Applying the same frontend-shift idea to dynamic surround-view reconstruction is a natural next step, since the paper explicitly leaves scene dynamics out of scope."],"forward_implications":["Freezing a geometry foundation and adapting only the neck and decoder is enough to reach state-of-the-art single-frame surround-view rendering, so future systems can avoid training large reconstruction backbones from scratch.","Transient warmup on coarse depth outperforms both no warmup and persistent depth supervision, indicating that scale anchoring matters mainly during early learning.","Decoupling geometry and appearance features improves both rendering quality and depth consistency, whereas fusing them into a single direct Gaussian head trades one for the other.","The same decoder receives better inputs from the neck; the reported gains are attributable to the frontend and the adaptation, not to a more expressive Gaussian head."],"supporting_citations":[{"why":"Supplies the frozen VGGT geometry foundation tokens that carry the transferable multi-view structure the method builds on.","marker":"[12]"},{"why":"Provides the Omni-Scene benchmark protocol, the pixel–volume decoder baseline, and the primary method it must beat.","marker":"[1]"},{"why":"Generates the coarse metric depth pseudo-labels used in Scale Warmup and rendered-depth supervision.","marker":"[11]"},{"why":"Produces the pseudo relative depth maps used in the PCC geometric-consistency evaluation.","marker":"[10]"},{"why":"Supplies the nuScenes surround-view driving dataset and the ego-pose evaluation protocol.","marker":"[5]"},{"why":"Serves as the strongest driving-centric feed-forward baseline and the direct comparison for the headline gains.","marker":"[2]"}],"fun_headline_variants":["Geometry priors front-load driving scene rendering","Pretrained VGGT powers surround-view Gaussian splatting","Frontend geometry beats decoder-heavy driving reconstruction","Frozen VGGT boosts sparse surround-view rendering quality","Shifting geometry to frontend lifts driving scene synthesis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result stands on the assumption that the pseudo-depth maps used for warmup supervision and for evaluating geometric consistency (from Metric3D-v2 and Depth Anything V2) are unbiased proxies for true metric 3D structure in low-overlap driving scenes.","fun_headline_variants_meta":{"raw":{"variants":["Geometry priors front-load driving scene rendering","Pretrained VGGT powers surround-view Gaussian splatting","Frontend geometry beats decoder-heavy driving reconstruction","Frozen VGGT boosts sparse surround-view rendering quality","Shifting geometry to frontend lifts driving scene synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000133,"raw_usage":{"total_tokens":1154,"prompt_tokens":979,"completion_tokens":175,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":100}},"tokens_in":595,"tokens_out":175,"duration_ms":2890,"temperature":1.0,"reasoning_tokens":100,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:30:22.255877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate rendered depth consistency against LiDAR point clouds or multi-view triangulated ground truth on nuScenes instead of pseudo relative depth. If VGGD's PCC advantage over Omni-Scene shrinks or reverses under true metric evaluation, or if removing Scale Warmup no longer hurts when trained with LiDAR labels, the paper's geometric-consistency claim is not supported.","supporting_citations":[{"cited_title":"Vggt: Visual geometry grounded transformer,","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen VGGT geometry foundation tokens that carry the transferable multi-view structure the method builds on."},{"cited_title":"Omni-scene: Omni-gaussian representation for ego-centric sparse-view scene reconstruction,","cited_arxiv_id":null,"evidence_quote":"Provides the Omni-Scene benchmark protocol, the pixel–volume decoder baseline, and the primary method it must beat."},{"cited_title":"Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation,","cited_arxiv_id":null,"evidence_quote":"Generates the coarse metric depth pseudo-labels used in Scale Warmup and rendered-depth supervision."},{"cited_title":"Depth anything v2,","cited_arxiv_id":null,"evidence_quote":"Produces the pseudo relative depth maps used in the PCC geometric-consistency evaluation."},{"cited_title":"nuscenes: A multimodal dataset for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Supplies the nuScenes surround-view driving dataset and the ego-pose evaluation protocol."},{"cited_title":"Drivingforward: Feed-forward 3d gaussian splatting for driving scene reconstruction from flexible surround- view input,","cited_arxiv_id":null,"evidence_quote":"Serves as the strongest driving-centric feed-forward baseline and the direct comparison for the headline gains."}],"review_version":1}