{"id":"6199d346-3b11-41eb-a4bf-15c7b82dddff","arxiv_id":"2507.12095","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"BRUM improves sparse-view 3D vehicle reconstruction by synthesizing extra training views from depth and poses, using DUSt3R for real cameras, and weighting the photometric loss by pixel reliability.","lead":"This paper presents BRUM, a method that reconstructs 3D vehicles from as few as 4 to 8 images by generating extra camera views with depth maps and pose interpolation, then training a Gaussian splatting model with a selective loss. It also introduces a new bus dataset and reports better reconstruction quality than three existing methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim is unsubstantiated: recent sparse-view Gaussian splatting baselines (CoherentGS, FewViewGS, SC-GS) cited in the paper are never compared, so gains over 3DGS/SplatFields do not establish SOTA.","rationale":"The reader's formal weakest_assumption is the accuracy of DUSt3R depth and pose estimates, but the more load-bearing gap for the central claim is the incomplete baseline set. The strongest claim is 'state-of-the-art performance', which is an empirical comparative statement; a comparison set lacking the current sparse-view GS methods does not substantiate it. The reader's rationale does mention missing baselines as a red flag, so there is partial agreement, but the emphasis differs. The DUSt3R dependency, while real, would only show that the method's real-world gains are conditional on a preprocessing model; if BRUM beats all strong baselines on the reported benchmarks, it could still be called SOTA on those benchmarks. Hyperparameter tuning on the test scenes is a serious protocol weakness, but it inflates absolute numbers rather than reversing the relative ordering unless the baselines were also tuned, which the paper does not report. The background-masking mechanism in Eq. 6-9 also contains an internal inconsistency: the text claims background pixels are retained for optimization, but the weighting in Eq. 7-8 assigns zero weight to pixels with no projected points, effectively zeroing out background contributions; this suggests the described XNOR operation may not behave as claimed. This is not the primary threat to SOTA, but it indicates the method's description needs closer scrutiny. Overall, the paper presents a plausible augmentation strategy and a useful dataset, but the empirical evidence is insufficient to support the unconditional 'state-of-the-art' label; a conditional verdict, contingent on the missing comparisons, is appropriate.","tokens_in":12570,"tokens_out":13492,"duration_ms":154755,"concrete_test":"Run the official implementations of CoherentGS, FewViewGS, and SC-GS on the released BRUM-dataset and CarPatch using the identical protocol: 4 training views with GT depth/poses for synthetic scenes, and 8 views with DUSt3R poses for real scenes, evaluating PSNR, SSIM, LPIPS, and AVGE. If any of these baselines matches or exceeds BRUM on the primary metrics, the state-of-the-art claim fails; if BRUM remains the best, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: BRUM achieves state-of-the-art performance with 4-8 images (abstract, Tables I, II). The experiments, however, compare only against 3DGS, DNGaussians, and SplatFields. The paper itself cites CoherentGS [36], FewViewGS [37], and SC-GS [38] in Related Work (Sec. II) as sparse-view Gaussian splatting methods, but they are absent from the evaluation. These methods are specifically designed for few-view inputs and typically report substantial improvements over 3DGS on standard benchmarks. Without them, the reported gains (e.g., +0.85 PSNR over SplatFields on BRUM-dataset synthetic, Table I) do not demonstrate state-of-the-art; a stronger omitted baseline could exceed BRUM. The DUSt3R dependency flagged by the reader is a secondary robustness concern: the real-world results rely on DUSt3R depth/pose quality, but this affects generalization rather than the core SOTA claim on the tested benchmarks. Hyperparameter tuning of c and h on the same scenes used for final results (Sec. IV-C, Tables III, IV) and the absence of variance estimates further weaken the quantitative evidence, but the missing baselines are the most direct threat to the headline assertion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BRUM, a method for sparse-view 3D vehicle reconstruction that augments Gaussian splatting training with synthetic views. For each input camera, BRUM generates novel poses by SLERP-interpolating toward the two nearest cameras, back-projects the input depth into a point cloud, renders synthetic views via point splatting, and trains 3DGS with a masked, confidence-weighted L1 loss on the generated views while retaining the standard loss on real views. For real-world scenes, DUSt3R replaces COLMAP for camera pose and depth estimation, with a confidence threshold and segmentation masks. The authors also introduce the BRUM-dataset with six synthetic and six real bus scenes, and report improvements over 3DGS, DNGaussians, and SplatFields on the BRUM-dataset, CarPatch, and KRONC.","tokens_in":12820,"tokens_out":3124,"duration_ms":37209,"significance":"If fully substantiated, the contribution is practically valuable: the method is conceptually simple, introduces a new dataset for an under-served domain, and ships code and data publicly. The ablation in Table V gives useful evidence that the masked XNOR operation, the confidence weighting, and the exclusion of SSIM on generated views each contribute. However, the central claim of state-of-the-art performance is currently supported only against three baselines, one of which (DNGaussians) is configured with defaults from an unrelated scene, and the reported numbers are obtained after tuning key hyperparameters on the evaluation datasets. The evidence is therefore not yet strong enough to support the abstract's claim as stated.","major_comments":[{"comment":"The state-of-the-art claim is not established because the authors cite CoherentGS [36], FewViewGS [37], and SC-GS [38] as sparse-view Gaussian splatting methods in Related Work but never compare against them in Tables I or II. These methods are specifically designed for few-view inputs and typically show large gains over plain 3DGS; a missing baseline could exceed BRUM. The authors should either add these or other recent sparse-view GS baselines to the comparison, or explicitly qualify the claim from 'state-of-the-art' to 'state-of-the-art among the compared methods'.","section":"Sec. II and Tables I-II"},{"comment":"Hyperparameters h and c are selected on the same data used for the final results. Table III tunes the SLERP range h on the Ford scene of KRONC and Table IV tunes the DUSt3R confidence threshold c on bus 1 of the BRUM-dataset; the chosen values are then applied to report the full-dataset numbers in Tables I and II. Because there is no validation split or cross-validation, the reported improvements may reflect test-set tuning. The authors should report results with fixed defaults or perform scene-wise cross-validation, and clarify whether the chosen values generalize to held-out scenes.","section":"Sec. IV-C, Tables III-IV"},{"comment":"The quantitative results are reported as dataset averages without standard deviations or per-scene breakdowns, and Figure 3 plots AVGE curves without error bars. With only six real and six synthetic scenes per dataset, the observed gaps (e.g., +0.85 PSNR over SplatFields on the synthetic BRUM-dataset in Table I) may be within scene-to-scene variability. The authors should report standard deviations, per-scene results, or significance tests so the reader can assess whether the differences are meaningful.","section":"Tables I-II and Fig. 3"},{"comment":"The real-world robustness claim depends on DUSt3R's depth and pose estimates, but the paper provides no direct evaluation of those estimates. The synthetic experiments use ground-truth depth and poses for augmentation (Sec. IV-B), while the real experiments use DUSt3R with a confidence threshold (Sec. III-C) and a reduced SLERP range; the effect of estimation error is not isolated. An ablation that replaces DUSt3R with COLMAP or uses oracle depth/pose on the real scenes would clarify whether BRUM is robust to the preprocessing errors it is designed to handle.","section":"Sec. III-C and Sec. IV-B vs IV-C"}],"minor_comments":[{"comment":"There is a typo: 'adpot' should be 'adopt'.","section":"Sec. III-C"},{"comment":"There is a typo: 'experimentis' should be 'experiments'.","section":"Sec. IV-C"},{"comment":"The symbol ⊕ is defined as XOR, but the operation 1−(V̂ ⊕ V̂) is XNOR; the text calls it 'exclusive nor' and Table V calls it 'XNOR'. Please use a single, explicit notation (e.g., ⊙ for XNOR or write the formula directly) to avoid confusion.","section":"Eq. (6) and Table V"},{"comment":"The description of h is ambiguous: the text says 'h is fixed at 0.1' while Table III lists h values of 0.05, 0.1, 0.3, 0.5. Clarify whether h denotes the maximum interpolation value or each sampled value, and how the step size 0.025 interacts with these numbers.","section":"Sec. IV-C, Table III"},{"comment":"Synthetic images are rendered at 800×800 while real images are downsampled to 512×256 for DUSt3R; please state whether the final evaluation was also performed at 512×256 or at the original resolution, as this affects the comparability of metrics across settings.","section":"Sec. IV-B and IV-C"}],"recommendation":"major_revision","confidential_remarks":"The missing-baseline issue is the most direct threat to the headline claim, and the test-set tuning of h and c is a serious methodological concern. CarPatch and KRONC come from the authors' prior work, which is not disqualifying but makes the absence of independent sparse-view baselines more salient. I would encourage the editor to ask for a comparison with CoherentGS, FewViewGS, or SC-GS, and for a validation protocol for h and c, before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead BRUM. The core idea is a sparse-view GS augmentation pipeline: use DUSt3R for pose and depth, SLERP to synthesize intermediate poses, splat points into those novel views, and train GS with a reliability-weighted, masked L1 loss so bad pixels don't pull the optimization. The components are known, but the combination is new, and the XNOR masking trick to keep background Gaussians alive during synthetic-view training is a genuinely neat detail. The BRUM-dataset (6 synthetic, 6 real buses, with masks and code release) is a concrete asset for vehicle reconstruction and inspection.\n\nWhat the paper does well: the method is described precisely enough to reproduce, the ablation in Table V supports each piece of the loss, and the qualitative results look plausible. The synthetic setup is honest about using GT depth and pose for augmentation, which is the right way to isolate the contribution of the augmentation itself.\n\nThe soft spots are mostly at the level of empirical support. The abstract says \"state-of-the-art performance,\" but the experiments compare only against 3DGS, DNGaussians, and SplatFields. CoherentGS, FewViewGS, and SC-GS appear in the Related Work as sparse-view GS methods and are never benchmarked. Those methods are designed for exactly the regime this paper targets; skipping them makes the SOTA claim unsupported. Also absent are variance estimates; a single run per scene leaves the reported margins (e.g., +0.85 PSNR over SplatFields) ungrounded. The hyperparameter analysis for h and c is done on the Ford scene (KRONC) and bus 1 (BRUM) and then applied to dataset-wide averages, with no separate validation split. That is a mild leak, not fatal, but it should be fixed. The reliance on DUSt3R for real-world poses is a real robustness concern, but it is a generalization limitation rather than a problem with the in-dataset numbers. Note also that CarPatch and KRONC come from the authors' own prior work; not disqualifying, but it makes the missing baselines more conspicuous.\n\nWho this is for: people working on sparse-view reconstruction of vehicles, automated inspection, or fleet monitoring. It deserves a serious referee. I would send it to review, not desk-reject, but I would ask for the missing sparse-view baselines, standard deviations, and a clean validation split for h and c before accepting.\n\nRegards.","headline":"Useful dataset and a clean augmentation trick; the 'state-of-the-art' claim is not yet supported by the comparisons.","tokens_in":13392,"tokens_out":3514,"would_cite":true,"duration_ms":38069,"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":"From as few as four 360° images of a vehicle, BRUM reconstructs the full 3D shape by synthesizing views along interpolated camera paths and training Gaussian splatting only on high-confidence pixels.","keywords":["3D vehicle reconstruction","sparse-view novel view synthesis","Gaussian splatting","SLERP view augmentation","DUSt3R pose estimation","confidence-weighted photometric loss","BRUM-dataset","360-degree capture"],"falsifier":"Take a real bus scene, run BRUM with DUSt3R for poses and depth, then rerun it with poses from a surveyed rig and depth from LiDAR or structure-from-motion on dense video while keeping every other setting the same; if the reconstruction quality stays close to the DUSt3R version, the gains are robust to geometry errors, and if it jumps, the reported real-world advantage rests on DUSt3R's accuracy rather than on the augmentation and loss design.","tokens_in":12364,"feed_emoji":"🚌","tokens_out":6367,"duration_ms":68205,"temperature":0.7,"pith_summary":"This paper tries to show that accurate 3D vehicle models can be built from as few as four to eight photos arranged around a vehicle, a setting where standard radiance-field and Gaussian-splatting pipelines fail or need dense captures. The proposed method, BRUM, creates extra training views by interpolating camera poses with SLERP and projecting each input image's depth map into those synthetic poses, then trains a Gaussian-splatting model with a loss that ignores pixels that were not reliably re-projected. For real-world captures it replaces the usual structure-from-motion pose estimator with DUSt3R, whose depth and pose estimates are filtered by a confidence threshold. If correct, the method makes vehicle inspection, predictive maintenance, and fleet monitoring practical with quick sparse photo captures instead of dense video. A new public dataset of six synthetic and six real buses is introduced to support this evaluation.","feed_headline":"Four bus photos yield a full 3D model via synthesized in-between views","feed_subtitle":"A sparse-view Gaussian splatting pipeline reaches state-of-the-art quality on car and bus benchmarks.","key_machinery":"The load-bearing mechanism is a geometric view-augmentation loop wrapped around Gaussian splatting. For each input camera, SLERP interpolates its pose with its two nearest cameras over small rotation ranges to sample M synthetic poses; the image and depth map are back-projected into a point cloud and rendered into those poses by splatting 3D points onto disks (Eq. 4). An exclusive-NOR mask (Eq. 6) keeps background pixels while removing pixels that did not reproject, and a per-pixel weight (Eqs. 7-8) derived from the splat influence rewards pixels receiving many consistent 3D points. Synthetic views train the Gaussians with a masked, weighted L1 loss only, dropping SSIM because the generated images lack reliable global structure. For real scenes, DUSt3R supplies poses, depth, and per-point confidence, and Eq. 10 prunes the point cloud by a confidence threshold c.","core_discovery":"The central claim is that sparse-view 360-degree vehicle reconstruction can be made practical by view augmentation plus selective supervision, rather than by a more complex network. With four input views in synthetic scenes and eight in real scenes, BRUM reports the best numbers on the new BRUM-dataset, on CarPatch, and on KRONC across PSNR, SSIM, LPIPS, and AVGE; for example it reaches 25.57 PSNR and 0.040 AVGE on CarPatch under four views, against 23.39 PSNR and 0.056 AVGE for the strongest compared baseline. The improvement comes from giving Gaussian splatting many synthetic views that lie between the real captures, while shielding the optimization from the unreliable pixels those synthetic views contain.","pith_inferences":["Inference: if DUSt3R's depth or pose errors are the limiting factor, coupling the pipeline with pose and depth refinement during Gaussian optimization could push real-world scores closer to the synthetic ones.","Inference: the same SLERP-plus-masked-loss recipe may transfer to other object-centric sparse-view tasks, such as furniture, industrial parts, or people, wherever depth and poses can be obtained.","Inference: the confidence threshold c is dataset-specific in the paper, so an automatic per-scene selection of c, or a soft weighting based on DUSt3R confidence, is a natural next step.","Inference: because the synthetic benchmark uses ground-truth depth and poses, the real-world results should be read as an upper bound on what the augmentation can add when geometry estimation is perfect."],"forward_implications":["With only 4-8 images per vehicle, a capture session becomes quick enough for daily inspection rounds and fixed multi-camera setups.","Generated views act as free training data, so the method avoids the architectural overhead and extra inference cost of learned sparse-view priors.","The selective weighted loss prevents synthetic-view artifacts from corrupting the Gaussian optimization, which is what lets the augmentation help instead of hurt.","On denser 12-view inputs the gap to baselines narrows, indicating the method is most valuable precisely in the sparse regime it targets.","The released BRUM-dataset gives the vehicle-reconstruction community synthetic and real bus scenes with masks for evaluating sparse-view methods."],"supporting_citations":[{"why":"Supplies the Gaussian splatting representation and optimization that BRUM augments with synthetic views.","marker":"[2]"},{"why":"DUSt3R provides the camera poses, depth maps, point clouds, and per-point confidence used for real-world preprocessing.","marker":"[12]"},{"why":"SLERP defines the quaternion interpolation that generates novel camera poses along geodesic arcs.","marker":"[39]"},{"why":"Provides the splat-based projection and z-buffer compositing used to render novel views from point clouds.","marker":"[40]"},{"why":"Segment Anything produces the foreground masks that focus reconstruction on the vehicle and filter the point cloud.","marker":"[44]"},{"why":"CarPatch is one of the two synthetic benchmarks used to evaluate sparse-view reconstruction quality.","marker":"[3]"},{"why":"KRONC is the real-world car benchmark used for evaluation and is the prior work this method extends.","marker":"[4]"},{"why":"DNGaussians is a sparse-view Gaussian splatting baseline that BRUM compares against.","marker":"[45]"},{"why":"SplatFields is a state-of-the-art sparse-view Gaussian splatting baseline in the comparisons.","marker":"[35]"}],"fun_headline_variants":["Sparse bus views build full 3D via synthesized in-betweens","Four views to 3D: vehicle model via synthesized views","3D vehicles from sparse views with view synthesis","Robust 3D vehicles from 360 sparse views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes the depth map and camera pose for each input view are accurate enough that back-projecting and re-rendering produces trustworthy synthetic views; in real scenes that accuracy comes from DUSt3R, so if its depth or pose errors are large the synthetic views feed wrong geometry into the Gaussian optimization.","fun_headline_variants_meta":{"raw":{"variants":["Sparse bus views build full 3D via synthesized in-betweens","Four views to 3D: vehicle model via synthesized views","3D vehicles from sparse views with view synthesis","Robust 3D vehicles from 360 sparse views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000155,"raw_usage":{"total_tokens":1173,"prompt_tokens":861,"completion_tokens":312,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":241}},"tokens_in":477,"tokens_out":312,"duration_ms":3887,"temperature":1.0,"reasoning_tokens":241,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:53:58.858751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real bus scene, run BRUM with DUSt3R for poses and depth, then rerun it with poses from a surveyed rig and depth from LiDAR or structure-from-motion on dense video while keeping every other setting the same; if the reconstruction quality stays close to the DUSt3R version, the gains are robust to geometry errors, and if it jumps, the reported real-world advantage rests on DUSt3R's accuracy rather than on the augmentation and loss design.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian splatting representation and optimization that BRUM augments with synthetic views."},{"cited_title":"Dust3r: Geometric 3d vision made easy,","cited_arxiv_id":null,"evidence_quote":"DUSt3R provides the camera poses, depth maps, point clouds, and per-point confidence used for real-world preprocessing."},{"cited_title":"Animating rotation with quaternion curves,","cited_arxiv_id":null,"evidence_quote":"SLERP defines the quaternion interpolation that generates novel camera poses along geodesic arcs."},{"cited_title":"Synsin: End-to- end view synthesis from a single image,","cited_arxiv_id":null,"evidence_quote":"Provides the splat-based projection and z-buffer compositing used to render novel views from point clouds."},{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Segment Anything produces the foreground masks that focus reconstruction on the vehicle and filter the point cloud."},{"cited_title":"Carpatch: A synthetic benchmark for radiance field evaluation on vehicle components,","cited_arxiv_id":null,"evidence_quote":"CarPatch is one of the two synthetic benchmarks used to evaluate sparse-view reconstruction quality."},{"cited_title":"Kronc: Keypoint-based robust camera optimization for 3d car reconstruction,","cited_arxiv_id":null,"evidence_quote":"KRONC is the real-world car benchmark used for evaluation and is the prior work this method extends."},{"cited_title":"Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization,","cited_arxiv_id":null,"evidence_quote":"DNGaussians is a sparse-view Gaussian splatting baseline that BRUM compares against."},{"cited_title":"Splatfields: Neural gaussian splats for sparse 3d and 4d reconstruction,","cited_arxiv_id":null,"evidence_quote":"SplatFields is a state-of-the-art sparse-view Gaussian splatting baseline in the comparisons."}],"review_version":1}