{"id":"6e796eac-1abb-4bdb-a4f8-d878186ec0cf","arxiv_id":"2501.02363","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"V2X-DGPE combines knowledge distillation, feature compensation, temporal fusion, and deformable attention to achieve state-of-the-art collaborative 3D detection on DAIR-V2X.","lead":"This paper proposes V2X-DGPE, a framework for vehicle-infrastructure collaborative 3D object detection that handles sensor domain differences and GPS pose noise. On the DAIR-V2X benchmark it reports the best accuracy so far and improved robustness to noise.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline SOTA margin is fully attributable to temporal fusion: Table III shows V2X-DGPE without Temporal Fusion (0.783/0.662) is no better than DI-V2X (0.788/0.662), and adding Temporal Fusion (+1.8/+3.3) exceeds the claimed 1.1/3.3 gain.","rationale":"The reader's weakest assumption identifies the same load-bearing concern, and the manuscript's own ablation strengthens it into a nearly decisive point. Table III shows that removing Temporal Fusion leaves V2X-DGPE at 0.783/0.662, essentially tying or slightly trailing DI-V2X (0.788/0.662), while adding Temporal Fusion produces the full headline gain. Since no baseline in Table I, II, or IV conditions on historical BEV frames, the comparison is unfair and the SOTA claim is not yet supported as an achievement of the domain-gap/pose-error modules. This is not an internal inconsistency; it is an uncontrolled comparison that a straightforward temporal-fusion baseline would settle. The released code makes such a test feasible. I do not see a more load-bearing concern: the pose-noise comparisons share the same confound, and the missing teacher specification and absent error bars are secondary. The reader's CONDITIONAL verdict remains appropriate: the empirical claim may survive, but only after controlling for temporal information.","tokens_in":10767,"tokens_out":3577,"duration_ms":34316,"concrete_test":"Re-run the DAIR-V2X experiment with a DI-V2X baseline augmented by the same Temporal Fusion Module (same historical frame size, same affine warp, same residual fusion), keeping training and evaluation protocol identical. If DI-V2X+TF reaches or exceeds 0.797/0.684 at AP@0.5/0.7, the SOTA claim is explained by temporal information and not by V2X-DGPE's domain/pose modules. If DI-V2X+TF cannot match, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—V2X-DGPE outperforms DI-V2X by 1.1%/3.3% AP@0.5/0.7 (Table I)—is not established as a property of the proposed domain-gap and pose-error modules. Table III's ablation isolates the Temporal Fusion Module: the pipeline with Knowledge Distillation, Feature Compensation, and Collaborative Fusion (no temporal) reaches 0.783/0.662, which is below DI-V2X at AP@0.5 (0.788) and equal at AP@0.7 (0.662). Adding Temporal Fusion gives 0.797/0.684, contributing +1.8/+3.3. The entire headline margin (+1.1/+3.3) is therefore accounted for by historical-BEV input. No Table I baseline consumes historical frames, so the comparison is not controlled for temporal information. Tables II and IV have the same confound: 'Ours' includes Temporal Fusion while DI-V2X does not, so even the pose-robustness margins may reflect extra input rather than deformable attention. The paper needs a temporal-fusion-equipped baseline to support the SOTA claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes V2X-DGPE, a feature-level collaborative 3D object detection framework for vehicle-infrastructure cooperation. It combines knowledge distillation, a feature compensation module, temporal fusion of historical BEV features, and a collaborative fusion module that uses heterogeneous self-attention and deformable attention to handle domain gaps and pose errors. Experiments on the real-world DAIR-V2X dataset report state-of-the-art results, outperforming DI-V2X by 1.1%/3.3% AP@0.5/0.7, and claim robustness to Gaussian and Laplace pose noise.","tokens_in":11002,"tokens_out":1967,"duration_ms":19223,"significance":"If the reported results are established under controlled comparisons, the architecture would be a meaningful step in V2X collaborative perception, since DAIR-V2X is a real-world benchmark and the paper releases code. The proposed modules are reasonable and the ablation study is informative. However, the central SOTA claim is not currently supported because the only comparison that isolates the proposed domain-gap and pose-error modules shows no improvement over DI-V2X; the entire headline gain is accounted for by the temporal fusion input, which is absent from all baselines. This is a load-bearing issue that must be resolved before the contribution can be assessed.","major_comments":[{"comment":"The headline claim that V2X-DGPE outperforms DI-V2X by 1.1%/3.3% (Table I) is not established as a property of the proposed domain-gap and pose-error modules. Table III shows that the full pipeline without Temporal Fusion (Knowledge Distillation + Feature Compensation + Collaborative Fusion) reaches 0.783/0.662, which is below DI-V2X at AP@0.5 (0.788) and equal at AP@0.7 (0.662). Adding Temporal Fusion gives 0.797/0.684, contributing +1.8/+3.3, which fully accounts for the reported margin. Since no Table I baseline consumes historical BEV frames, the comparison is confounded by extra input information. Please provide a baseline equipped with the same temporal-fusion input, or report the SOTA comparison with the temporal fusion module removed, so the contribution of the proposed modules can be isolated.","section":"Table III and Section IV-C"},{"comment":"The pose-robustness comparisons in Tables II and IV have the same confound as Table I: the row labeled 'Ours' includes the Temporal Fusion Module, while all baselines, including DI-V2X, do not. The improvements under Gaussian and Laplace noise may therefore reflect additional historical input rather than the deformable attention mechanism or other proposed modules. Please report the pose-noise results for the proposed method without temporal fusion, or add a temporal-fusion-equipped variant of DI-V2X as a baseline.","section":"Tables II and IV"},{"comment":"The knowledge distillation teacher is never specified. The paper states that the student features Bf are aligned with teacher BEV features Bt, but does not state what model produces Bt, what input (vehicle, infrastructure, or both) is used, how the teacher is trained, or which loss and alignment procedure are applied. Without this information, the knowledge distillation component is not reproducible and its contribution cannot be evaluated. Please specify the teacher architecture, training protocol, and alignment loss.","section":"Section III-A and experimental setup (Section IV-B)"},{"comment":"All tables report a single run with no standard deviations, multiple seeds, or statistical significance tests. The reported margins are small (for example, 0.797 vs 0.788 at AP@0.5 in Table I), so it is unclear whether the differences are stable or within run-to-run variation. Please provide results over multiple seeds with mean and standard deviation, or otherwise justify that the differences are not noise.","section":"Tables I-IV"}],"minor_comments":[{"comment":"The section heading 'Pose Errors Reslut' contains a typo; it should read 'Pose Errors Result'.","section":"Section IV-E"},{"comment":"The related-work paragraph contains an unresolved citation placeholder '[ ?]' for intermediate fusion; please replace it with the intended reference.","section":"Section II"},{"comment":"The sentence 'the learning rate decayes steadily' contains a typo; 'decayes' should be 'decays'.","section":"Section IV-B"},{"comment":"The 'Publication' column lists venue names without years for some entries (e.g., 'CVPR 2019') and 'AAAI 2024' for DI-V2X; please make the format consistent across all rows.","section":"Table I"},{"comment":"The text says the infrastructure BEV features are sent to the vehicle and then input into the Temporal Fusion Module, but Figure 2 is not described in enough detail to confirm which branch the historical features come from and whether the temporal module is applied to both agents; please clarify the data flow in the figure caption or text.","section":"Figure 2 and Section III-A"},{"comment":"The temporal fusion description mentions 'all detected objects features' after motion compensation, but no detection occurs before fusion; this phrasing is confusing and should be reworded to refer to BEV features or object queries.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The paper is in-scope for a computer vision / intelligent vehicles venue and the code release is a plus. The central empirical claim is currently confounded by the temporal-fusion input, and the missing teacher specification is a reproducibility concern. These are fixable with additional controlled experiments and clarifications, so I do not recommend rejection, but the revision needs to address the confound directly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's headline claim does not survive its own ablation. Table III shows the pipeline with KD, Feature Compensation, and Collaborative Fusion but no Temporal Fusion at 0.783/0.662, versus DI-V2X at 0.788/0.662. Adding Temporal Fusion gives 0.797/0.684, so the entire reported advantage over DI-V2X comes from the extra historical-BEV input. No baseline in Table I or the robustness tables consumes historical frames, so the comparison is not controlled for temporal information. The stress-test note is right, and it is the central issue.\n\nWhat is new: the specific integration of temporal fusion with the known building blocks (heterogeneous self-attention from V2X-ViT, deformable attention from Deformable DETR, knowledge distillation from DI-V2X) is legitimate extension work, evaluated on real-world DAIR-V2X with code released. The ablation is actually refreshingly honest because it lets you see the confound clearly.\n\nThe other soft spots are minor by comparison but real: no variance or multiple seeds, the knowledge-distillation teacher is never specified, and the paper says it beats DI-V2X by 1.1%/3.3% while Table I shows 0.9/2.2 percentage points difference. That may be a rounding or relative-percentage choice, but it needs stating.\n\nThis paper is for people working on practical V2X collaborative perception who want a feature-level fusion recipe with temporal modeling. The method is a reasonable extension of DI-V2X and V2X-ViT, and the code makes it reproducible. But the claim that the proposed domain-gap and pose-error modules jointly achieve SOTA is not supported; the evidence supports \"temporal fusion helps.\"\n\nSend it to peer review. A serious referee can require the controlled baseline, and with that baseline the paper could be solid. As it stands, re-frame the contribution and add the missing experiments before acceptance.","headline":"The claimed SOTA margin is entirely due to the temporal fusion input, which the comparisons never control for; without it the method matches, not beats, DI-V2X.","tokens_in":11525,"tokens_out":3404,"would_cite":false,"duration_ms":32654,"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":"V2X-DGPE claims state-of-the-art collaborative 3D detection on DAIR-V2X by combining knowledge distillation, feature compensation, temporal fusion, and deformable attention to close domain gaps and absorb pose errors.","keywords":["V2X collaborative perception","3D object detection","domain gap","pose error","knowledge distillation","deformable attention","temporal fusion","DAIR-V2X"],"falsifier":"Train DI-V2X extended with the same Temporal Fusion Module and historical BEV input on DAIR-V2X. If that baseline reaches or exceeds V2X-DGPE's 0.797/0.684, the improvement attributed to the domain-gap and pose-error components is not supported, especially since the paper's own ablation shows that removing temporal fusion drops V2X-DGPE to roughly DI-V2X's level (0.783/0.662 vs 0.788/0.662).","tokens_in":10587,"feed_emoji":"🚗","tokens_out":6048,"duration_ms":52308,"temperature":0.7,"pith_summary":"This paper tries to establish that feature-level fusion between a vehicle and roadside infrastructure can be made accurate despite two real failure modes: the domain gap between different LiDAR sensors, and pose errors from GPS noise and latency. It claims its architecture, V2X-DGPE, learns domain-invariant features through teacher-student distillation, compensates for distribution differences before fusion, brings in historical BEV frames, and uses deformable attention to adaptively sample around shifted features. On the real-world DAIR-V2X dataset it reports state-of-the-art AP@0.5/0.7 of 0.797/0.684, improving on DI-V2X by 1.1%/3.3%, and it keeps the best numbers under Gaussian and Laplace pose noise. A sympathetic reader would care because robustness to pose errors and sensor heterogeneity is a precondition for deploying collaborative perception in real traffic.","feed_headline":"V2X-DGPE beats DI-V2X and holds up under pose noise","feed_subtitle":"Collaborative 3D detector beats previous best and degrades gracefully when GPS and latency introduce errors.","key_machinery":"The central machinery is a teacher-student Knowledge Distillation Framework in which the student learns domain-invariant representations of vehicle and infrastructure features. It contains a Feature Compensation Module of three weighted residual blocks using depthwise separable convolutions to narrow the vehicle-infrastructure BEV feature distribution gap before fusion; a Temporal Fusion Module that affine-warps historical BEV features into the current frame and fuses them through a residual network; and a Collaborative Fusion Module that combines heterogeneous multi-head self-attention with relation-specific weights for each agent pair and deformable attention, which learns query-agnostic sampling-point offsets to absorb pose-induced feature shifts.","core_discovery":"V2X-DGPE claims that feature-level fusion between vehicles and roadside infrastructure can be made robust to domain gaps and pose errors by aligning features only after fusion, compensating for distribution shifts before fusion, injecting historical BEV frames, and using heterogeneous and deformable attention. On DAIR-V2X it reports AP@0.5/0.7 of 0.797/0.684, beating the previous best DI-V2X by 1.1/3.3 points, and it maintains the best AP at all tested Gaussian and Laplace noise levels, including the strongest noise tested with σt=0.6m and σr=0.6°.","pith_inferences":["Because no baseline in the comparison tables is given historical BEV frames as input, the head-to-head margin over DI-V2X probably overstates the contribution of the domain-compensation and pose-adaptation modules; a temporal-fusion-equipped DI-V2X would be a cleaner comparator.","The deformable-attention offset mechanism is a generic fix for coordinate misalignment and could improve single-vehicle 3D detectors under localization noise, not only collaborative fusion systems.","The paper's Laplace-noise evaluation is a useful stress test; a further stress test would be sensor-type mismatch beyond LiDAR line count, such as camera-LiDAR heterogeneity or weather-induced domain shift."],"forward_implications":["If the results hold, V2X-DGPE sets a new state of the art on the real-world DAIR-V2X benchmark, with the larger gain at the stricter IoU 0.7 threshold where alignment quality matters most.","Under Gaussian pose noise with σt=0.6m and σr=0.6°, it reports AP@0.7 of 0.594 versus 0.577 for DI-V2X; under Laplace noise it reports 0.588 versus 0.569, indicating more graceful degradation in both noise regimes.","The ablation study attributes 1.8 AP@0.5 and 3.3 AP@0.7 to the Temporal Fusion Module, suggesting that historical BEV context is an inexpensive accuracy lever for collaborative perception.","The deformable attention component offers a generic way to handle coordinate misalignment, so the architecture could transfer to other multi-agent perception settings with heterogeneous sensors and localization noise."],"supporting_citations":[{"why":"Supplies the DAIR-V2X real-world dataset used for all detection and pose-noise experiments.","marker":"[28]"},{"why":"DI-V2X is the state-of-the-art baseline that V2X-DGPE outperforms and the knowledge-distillation approach it directly extends.","marker":"[12]"},{"why":"V2X-ViT contributes the heterogeneous multi-agent self-attention design that the Collaborative Fusion Module builds on.","marker":"[26]"},{"why":"Deformable DETR is the source of the deformable attention mechanism used to sample around pose-shifted features.","marker":"[31]"},{"why":"PointPillars is the BEV feature extraction backbone shared by the student and teacher models.","marker":"[10]"},{"why":"Spatial transformer networks provide the affine grid sampling method used in the Temporal Fusion Module to align historical features.","marker":"[9]"},{"why":"DiscoNet supplies the teacher-student knowledge distillation paradigm for collaborative perception that V2X-DGPE adapts.","marker":"[14]"},{"why":"CoAlign is the strongest pose-robust collaborative baseline compared in the noise experiments.","marker":"[17]"}],"fun_headline_variants":["V2X-DGPE: Collaborative 3D detection robust to domain gaps and pose errors","V2X-DGPE beats DI-V2X on DAIR-V2X, resilient to GPS and latency noise","V2X-DGPE: Align features after fusion, inject history, and beat prior SOTA","V2X-DGPE uses deformable attention to counter pose noise; wins on DAIR-V2X"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume the comparison is apples-to-apples, but V2X-DGPE is the only evaluated method that consumes historical BEV frames as extra input, so the 1.1%/3.3% improvement over DI-V2X may come from the extra temporal information rather than from the proposed domain-gap and pose-error modules.","fun_headline_variants_meta":{"raw":{"variants":["V2X-DGPE: Collaborative 3D detection robust to domain gaps and pose errors","V2X-DGPE beats DI-V2X on DAIR-V2X, resilient to GPS and latency noise","V2X-DGPE: Align features after fusion, inject history, and beat prior SOTA","V2X-DGPE uses deformable attention to counter pose noise; wins on DAIR-V2X"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00089,"raw_usage":{"total_tokens":3817,"prompt_tokens":902,"completion_tokens":2915,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":2806}},"tokens_in":518,"tokens_out":2915,"duration_ms":18498,"temperature":1.0,"reasoning_tokens":2806,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:13:02.814841+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DI-V2X extended with the same Temporal Fusion Module and historical BEV input on DAIR-V2X. If that baseline reaches or exceeds V2X-DGPE's 0.797/0.684, the improvement attributed to the domain-gap and pose-error components is not supported, especially since the paper's own ablation shows that removing temporal fusion drops V2X-DGPE to roughly DI-V2X's level (0.783/0.662 vs 0.788/0.662).","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DAIR-V2X real-world dataset used for all detection and pose-noise experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DI-V2X is the state-of-the-art baseline that V2X-DGPE outperforms and the knowledge-distillation approach it directly extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"V2X-ViT contributes the heterogeneous multi-agent self-attention design that the Collaborative Fusion Module builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PointPillars is the BEV feature extraction backbone shared by the student and teacher models."},{"cited_title":"Jaderberg, K","cited_arxiv_id":null,"evidence_quote":"Spatial transformer networks provide the affine grid sampling method used in the Temporal Fusion Module to align historical features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DiscoNet supplies the teacher-student knowledge distillation paradigm for collaborative perception that V2X-DGPE adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CoAlign is the strongest pose-robust collaborative baseline compared in the noise experiments."}],"review_version":1}