{"id":"450201ca-9847-4b56-a35f-35c52cb98107","arxiv_id":"2608.12442","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MV2 is a multi-vehicle driving dataset with 50 scenes and 12,000 images that uses cross-vehicle and aerial-ground splits to show NVS quality falls sharply with viewpoint disparity.","lead":"A new driving dataset records the same urban scenes from a car, a scooter, and a drone with registered camera poses. It tests how well novel-view synthesis methods generalize to large viewpoint changes when trained on one vehicle and tested on another.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pose verification is self-consistency only: Section 3.5 admits absolute pose accuracy cannot be verified, and a <=30 px epipolar threshold on relative pairs does not rule out global SfM drift.","rationale":"The reader's weakest assumption and mine coincide: epipolar self-consistency is not ground truth. This is the most load-bearing concern because every downstream result—the PSNR rankings, the baseline-vs-drop trend, and the pose-estimation comparison—is computed in the released pose frame. If that frame is globally wrong in a way that passes the threshold, the whole benchmark is unverified. I considered the photometric confound (TC->L is a different camera) and the wall-clock synchronization issue; the TC->L control makes the photometric explanation less likely, since the drop to TC->L is small, and synchronization is at least partially checked. The pose issue is more fundamental and is explicitly conceded in Section 3.5. The paper has real strengths: the multi-platform capture, the cross-vehicle evaluation protocol, the depth-filtering ablation, and the pose-estimation comparison are valuable and reproducible in principle. No formal verification exists, but the dataset release and protocols are checkable. My recommendation is unchanged from the reader: CONDITIONAL, pending external pose validation. This is not a rejection; it is a precise condition that can be satisfied with a modest additional measurement.","tokens_in":13593,"tokens_out":8593,"duration_ms":85066,"concrete_test":"On ~10 MV2 sequences, collect reference trajectories with RTK-GPS/IMU on car, scooter, and drone, or surveyed checkerboard/AprilTag targets visible from all three platforms. Compute absolute trajectory error (ATE, in meters and degrees) of all released COLMAP train and test poses that passed the Section 3.5 threshold. If median ATE is small relative to the cross-vehicle baseline (e.g., <0.5 m translation and <1 degree rotation), the epipolar verification is adequate and the NVS trend stands. If median ATE is large (e.g., >1 m / 2 degrees) while epipolar errors remain below 30 px, the pose-validation premise fails; benchmark rankings and the extrapolation conclusion must be recomputed with corrected poses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"MV2's central scientific claim is that cross-vehicle splits measure extrapolation: NVS performance drops monotonically as camera baseline grows (Table 1, PVG 27.01->20.23 from TC->C to TC->S). This claim presupposes that the released camera poses are accurate enough to render and compare images at the correct viewpoints. The paper's only pose verification is epipolar self-consistency (Section 3.5): train poses are checked on consecutive edges plus a long-range edge, test poses against a single nearest training image, with maximum epipolar error <=30 px. The paper states: 'Since ground-truth poses are unavailable, absolute pose accuracy cannot be directly verified.' This is the load-bearing gap. Epipolar error is necessary but not sufficient: a smoothly drifting SfM reconstruction can keep every local relative pose epipolar-consistent while accumulating global error, and a wrongly localized test image can still satisfy the epipolar constraint with its nearest neighbor. Because NVS metrics compare renders from these poses to real test images, systematic pose error would directly depress PSNR/SSIM/LPIPS and could mimic the reported baseline-vs-degradation trend. The paper's own 'hypothesis' that locally accurate edges imply accurate absolute poses is not established. No external reference (GPS/IMU, surveyed markers, LiDAR) is used, so the benchmark cannot distinguish method failure from pose failure. This does not invalidate the dataset's potential, but it makes the headline conclusion conditional on an unverified quantity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MV2, a multi-view multi-vehicle driving dataset for novel view synthesis, with synchronized captures from a car, a scooter, and a drone. The authors describe a collection pipeline, SfM-based pose estimation, and a manual correspondence-based epipolar verification step that filters 200 recorded sequences down to 50 scenes. They define two evaluation setups (Eval-Car-Train and Eval-Drone-Train) and report NVS benchmarks for static, dynamic, and feed-forward methods, plus a comparison of COLMAP against feed-forward pose estimators. The headline findings are that NVS performance degrades as the train-test viewpoint disparity increases (e.g., PVG PSNR drops from 27.01 on T_C->C to 20.23 on T_C->S in Table 1) and that feed-forward pose estimators exhibit larger epipolar errors than COLMAP.","tokens_in":13884,"tokens_out":9140,"duration_ms":81279,"significance":"The dataset fills a genuine gap: existing driving NVS benchmarks sample train and test views from one trajectory, while MV2 provides co-registered ground-level and aerial captures of the same dynamic scenes. If the pose verification can be made rigorous, the benchmark would be a valuable resource for studying extrapolation rather than interpolation, and the paper's careful comparison of optimization-based and feed-forward methods, depth-supervision ablations, and dynamic-mask ablations would be useful to the community. The main results are plausible and the release of data and protocol is a strength. However, the paper's central quantitative claims currently rest on pose estimates whose absolute accuracy is explicitly unverified, and the dataset documentation contains inconsistencies that must be fixed before the benchmark can be reproduced.","major_comments":[{"comment":"Equation (1) defines the epipolar error as e = |x2^T F x1|, but this is the algebraic epipolar residual, not a pixel distance; the point-to-line distance requires normalization by the length of the epipolar line coefficients. As written, the thresholds em <= 30 used for sequence filtering and the histograms in Fig. 4 are not measured in pixels. Because this threshold is what removes 150 of 200 scenes, the pose-filtering protocol and every downstream conclusion that relies on it need to be re-evaluated with a correctly normalized error, or the implementation should be stated explicitly.","section":"Section 3.5, Eq. (1)"},{"comment":"The verification of camera poses is self-consistency only. The paper states that 'Since ground-truth poses are unavailable, absolute pose accuracy cannot be directly verified,' and the proposed checks—epipolar consistency on consecutive training edges, one long-range edge, and test-to-nearest-training pairs—cannot rule out global SfM drift or a consistently wrong alignment between the vehicle trajectories. A smoothly drifting reconstruction can satisfy all pairwise epipolar constraints while accumulating large absolute errors, and a mislocalized test image can still satisfy the epipolar constraint with its nearest neighbor. Because every NVS number in Table 1 is computed by rendering at these poses, systematic pose error would directly depress the metrics and could mimic or exaggerate the reported baseline-vs-degradation trend. I recommend validating a subset of scenes against an external reference (e.g., GPS/IMU, surveyed markers, or LiDAR alignment) or, failing that, explicitly reframing the benchmark as a relative/self-consistent evaluation and removing the strong 'pose accuracy' wording.","section":"Section 3.5"},{"comment":"The dataset composition is internally inconsistent. Section 3.1 defines only four camera streams (VL, VC, VS, VD), yet Section 3.2 reports '5 camera sequences each' and Section 3.3 mentions a test set T_C->D that was never defined. The abstract's 12,000 images also do not follow from the stated numbers: 50 retained scenes with 5 cameras and 100 frames per camera would yield 25,000 images even before excluding every 5th test frame. The paper must reconcile the number of cameras, define all test sets, and report per-split image counts so the benchmark can be reproduced.","section":"Sections 3.1–3.3"},{"comment":"The camera-pose-estimation comparison is evaluated with the same epipolar-error criterion that was used to filter COLMAP poses and select the scenes, so the comparison is not independent: COLMAP is favored by construction, since the 50 scenes were retained precisely because COLMAP's poses passed this check, while feed-forward methods are scored on the same quantity. This does not invalidate the observation that feed-forward methods produce larger epipolar residuals, but the claim that they 'lag behind' optimization-based methods should be presented as a relative statement about this metric, or verified with an external pose reference.","section":"Section 5.5, Fig. 5"}],"minor_comments":[{"comment":"Many cross-references are unresolved in the submitted text: 'Section??' appears in Sections 2, 3.3, 4, 5.2, 5.3, 5.4, and 5.5, and 'Fig.??' appears in Section 4; these need to be filled in before publication.","section":"Throughout"},{"comment":"Section 3.1 states that the resolution is 1080×1980 while the abstract uses 1080×1920; please correct the inconsistency.","section":"Section 3.1"},{"comment":"Section 3.1 says 'the set of images collected using these three cameras' after defining only two car cameras; the number of cameras per vehicle should be stated unambiguously.","section":"Section 3.1"},{"comment":"In Section 3.3, the sentence beginning 'The following pairs of test sets' is grammatically incomplete and refers to T_C->D, which is not introduced in the Eval-Car-Train setup; please revise.","section":"Section 3.3"},{"comment":"Figure 6 caption contains the typo 'BBoth'.","section":"Figure 6 caption"},{"comment":"The paper does not quantify the actual camera baselines between T_C->C, T_C->L, T_C->S, T_D->C, etc.; since the central trend is attributed to increasing baseline, reporting average baseline distances or another disparity measure would strengthen the claim.","section":"Section 3.3 / Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not yet in a publishable form due to unresolved placeholders and the sensor-count inconsistency, but the underlying dataset and benchmark idea are sound. The pose-accuracy issue is the key correctness risk; I would like to see external validation before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe real contribution is the dataset: synchronized car, scooty, and drone captures of the same dynamic scenes, registered into 50 scenes / 12,000 images, with train-on-one-vehicle/test-on-another protocols. That is genuinely new relative to KITTI/Waymo/nuScenes single-trajectory benchmarks and to concurrent multi-traversal work. The manual correspondence annotation for epipolar verification is useful diligence, and the benchmark spans static, dynamic, optimization-based, and feed-forward NVS methods plus pose estimators. Table 1's pattern—PVG drops from 27.01 to 20.23 PSNR from TC->C to TC->S—is consistent across methods and supports the point that cross-vehicle splits measure extrapolation.\n\nThe soft spot is the one the stress-test note flags: Section 3.5 admits absolute pose accuracy cannot be verified, and the epipolar check only certifies relative consistency. A smoothly drifting SfM reconstruction can pass the <=30 px threshold while being globally wrong, and a wrongly localized test image can still satisfy the constraint with its nearest neighbor. That does not invalidate the dataset, but it makes the headline conclusion conditional. If poses are off, the NVS metrics and the pose-estimation comparison could be depressed by pose error rather than method failure. I don't think that is likely enough to reject the paper, but it is a load-bearing gap.\n\nOther issues are more minor. Filtering from 200 to 50 scenes with a hand-chosen epipolar threshold is aggressive and could bias toward easy scenes, though the retained distribution is shown. The manuscript has multiple Section?? placeholders and missing supplementary statistics. The WOD cross-lane comparison is qualitative only, and the depth-supervision comparison uses different settings, so treat that table as suggestive. The pose-estimation section uses epipolar error as a proxy for pose accuracy, which is reasonable given the lack of ground truth but means the reported lag of feed-forward methods is likewise conditional.\n\nWho is this for? Anyone building or evaluating driving NVS methods, especially on aerial-to-ground or cross-trajectory generalization. The dataset and protocol are a real step beyond interpolation-only benchmarks.\n\nRecommendation: send it to peer review. The authors should be asked to add external pose validation (GPS/IMU or surveyed markers) on a subset, complete the supplementary placeholders, and discuss the 200->50 filtering more carefully. With those changes the claims would be solid.","headline":"Useful new cross-vehicle NVS benchmark, but the headline numbers rest on pose self-consistency rather than absolute pose ground truth; worth a serious referee if the authors commit to validating or softening that claim.","tokens_in":14431,"tokens_out":2646,"would_cite":true,"duration_ms":24368,"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":"MV2 is a multi-vehicle driving dataset that shows novel-view synthesis quality collapses when train and test views come from different vehicles (car, scooter, drone).","keywords":["novel view synthesis","driving dataset","multi-vehicle capture","aerial-to-ground","3D Gaussian Splatting","camera pose estimation","cross-vehicle generalization","wide-baseline views"],"falsifier":"Register a handful of MV2 scenes with surveyed markers or RTK-GPS ground-truth poses, then compare those true poses against the Structure-from-Motion poses that passed the epipolar threshold; if large global pose errors coexist with sub-30-pixel epipolar errors, the certification criterion is insufficient.","tokens_in":13446,"feed_emoji":"🚗","tokens_out":4536,"duration_ms":38346,"temperature":0.7,"pith_summary":"This paper introduces MV2, a driving dataset in which the same urban scenes are captured simultaneously from a car, a scooter, and a drone. The authors use these synchronized multi-vehicle streams to define a novel-view-synthesis benchmark where models are trained on one vehicle's camera and tested on another's, producing much larger viewpoint changes than the same-trajectory splits used by earlier driving benchmarks. They report that every evaluated method loses quality as the train-test camera baseline grows, and that optimization-based rendering approaches clearly outperform feed-forward ones in the wide-baseline regime. The paper argues that such cross-vehicle evaluation measures true extrapolation, which matters for photorealistic driving simulation.","feed_headline":"Wide-baseline driving views break novel-view synthesis models","feed_subtitle":"Cross-vehicle tests show rendering quality falls as camera baselines grow, exposing a gap for driving simulation.","key_machinery":"The load-bearing mechanism is the multi-vehicle capture setup: synchronized car, scooter, and drone cameras looking at the same scene, registered into one coordinate frame by Structure-from-Motion. Poses are certified not by ground truth but by epipolar consistency: human-annotated region correspondences, refined with a dense neural matcher, must keep the mean and maximum epipolar error of every relative pose at or below 30 pixels. That certification step decides which of 200 recorded sequences survive as the 50-scene benchmark, so it is the gatekeeper for every numerical comparison in the paper.","core_discovery":"The central claim is that current novel-view-synthesis models, including dynamic Gaussian-splatting methods that lead on standard splits, degrade sharply when asked to render a view from a different vehicle's trajectory, and that existing single-trajectory benchmarks cannot expose this failure because they only interpolate between nearby frames. On MV2, the best dynamic method drops from 27.01 to 20.23 PSNR when moving from the same-trajectory test split to the car-to-scooter split, and aerial-to-ground splits are harder still. The paper also finds that feed-forward pose estimators produce much larger epipolar errors than optimization-based localization under these wide baselines, so the dataset doubles as a pose-estimation stress test.","pith_inferences":["The epipolar-error threshold of 30 pixels could be tested directly: if a subset of scenes were re-certified with surveyed ground-truth poses, the threshold's sufficiency could be quantified rather than assumed.","The same benchmark could be extended to closed-loop evaluation, where a rendered scooter or drone view is fed to a perception model; the paper notes downstream uses but does not run that experiment.","A natural follow-up is to check whether training feed-forward splatting models on multi-vehicle, wide-baseline data reduces the gap, since the paper only evaluates existing models trained on standard data."],"forward_implications":["Benchmarks that sample train and test frames from one vehicle trajectory overestimate generalization; MV2-style cross-vehicle splits should be used to measure extrapolation.","Aerial-to-ground synthesis, where current methods score below 13 PSNR, is an open problem rather than a solved one.","Depth supervision from monocular estimators, when filtered to the 10–30 m range, improves dynamic-scene methods but does not close the wide-baseline gap.","Feed-forward pose estimators need to handle large viewpoint gaps before they can replace optimization-based localization in driving scenes.","NVS quality on cross-vehicle splits is a direct proxy for whether a simulator can render a plausible lane-change maneuver."],"supporting_citations":[{"why":"Supplies the optimization-based Structure-from-Motion pipeline used to register all camera poses.","marker":"[30]"},{"why":"Provides the leading dynamic Gaussian-splatting baseline whose performance drop defines the cross-vehicle gap.","marker":"[6]"},{"why":"Provides the static 3D Gaussian Splatting baseline used throughout the benchmark.","marker":"[17]"},{"why":"One of the feed-forward splatting baselines compared under wide-baseline test splits.","marker":"[5]"},{"why":"One of the feed-forward splatting baselines compared under wide-baseline test splits.","marker":"[40]"},{"why":"One of the feed-forward splatting baselines compared under wide-baseline test splits.","marker":"[23]"},{"why":"Provides the dense neural feature matcher used to obtain pixel correspondences for pose verification.","marker":"[9]"},{"why":"Provides a standard single-trajectory driving dataset used to show MV2's same-trajectory performance is comparable.","marker":"[31]"},{"why":"Supplies the off-the-shelf monocular depth estimator used as LiDAR-replacement supervision.","marker":"[21]"},{"why":"Supplies segmentation masks intersected with motion masks to obtain dynamic-object supervision.","marker":"[18]"}],"fun_headline_variants":["Cross-vehicle viewpoints shatter novel-view synthesis","Driving NVS fails when cameras span wide baselines","MV2: new stress test for dynamic driving renderers","Wide-baseline driving views reveal model blind spots","Pose estimators stumble on cross-vehicle driving data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark assumes that a maximum epipolar error of 30 pixels on a few dozen manually annotated correspondences is enough to certify that every camera pose is accurate, because no independent ground-truth pose source exists.","fun_headline_variants_meta":{"raw":{"variants":["Cross-vehicle viewpoints shatter novel-view synthesis","Driving NVS fails when cameras span wide baselines","MV2: new stress test for dynamic driving renderers","Wide-baseline driving views reveal model blind spots","Pose estimators stumble on cross-vehicle driving data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001009,"raw_usage":{"total_tokens":4235,"prompt_tokens":883,"completion_tokens":3352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":3274}},"tokens_in":499,"tokens_out":3352,"duration_ms":24845,"temperature":1.0,"reasoning_tokens":3274,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:11:43.522348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Register a handful of MV2 scenes with surveyed markers or RTK-GPS ground-truth poses, then compare those true poses against the Structure-from-Motion poses that passed the epipolar threshold; if large global pose errors coexist with sub-30-pixel epipolar errors, the certification criterion is insufficient.","supporting_citations":[{"cited_title":"In: ECCV (2024)","cited_arxiv_id":null,"evidence_quote":"One of the feed-forward splatting baselines compared under wide-baseline test splits."},{"cited_title":"In: CVPR (2024)","cited_arxiv_id":null,"evidence_quote":"One of the feed-forward splatting baselines compared under wide-baseline test splits."},{"cited_title":"In: CVPR (2025)","cited_arxiv_id":null,"evidence_quote":"One of the feed-forward splatting baselines compared under wide-baseline test splits."},{"cited_title":"In: CVPR (2020)","cited_arxiv_id":null,"evidence_quote":"Provides a standard single-trajectory driving dataset used to show MV2's same-trajectory performance is comparable."}],"review_version":1}