{"id":"f5c6ffc1-d1f0-4fb1-a2c4-e6a18e873430","arxiv_id":"2508.04702","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A contrastive learning framework with instance-level and perspective-level losses consistently improves multiple BEV detection models on nuScenes by up to 2.4 mAP.","lead":"BEVCon adds two contrastive learning training losses to bird's-eye-view camera perception models, improving 3D object detection on the nuScenes benchmark. The method is a general training recipe that boosts several BEV detectors without new data or labels, which matters for self-driving perception.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline +2.4% mAP gain for BEVFormer-tiny (0.276) conflicts with the paper's own repeated-run mean for the full framework (0.2735±0.0002, Tables III/IV), so the central quantitative claim is not internally reproducible as reported.","rationale":"The paper is a straightforward empirical contribution: two auxiliary contrastive losses added to BEV detectors improve nuScenes mAP. The method is described clearly, and the experiments cover three distinct detector families. The availability of code and the inclusion of repeated-run means in the ablation are positives. The reader's weakest assumption (BEV-space alignment under augmentation) is a plausible implementation risk, but the paper explicitly states that the augmentation process ensures alignment (Sec. IV.B) and the claim is checkable in the released code; it is not contradicted by anything in the manuscript. In contrast, the discrepancy between the headline tables and the ablation tables is directly observable and concerns the central quantitative claim itself. Table I lists BEVFormer-tiny + BEVCon as 0.375 NDS / 0.276 mAP; Table III's Contrast Framework (and Table IV(a) full row) gives 0.371 / 0.274, with std ≤0.0002. The difference is many times the reported standard deviation, so it cannot be attributed to seed noise. The same pattern appears for ResNet-101-DCN. This means the manuscript contains two mutually inconsistent sets of numbers for the same claimed configuration. A reader cannot tell which set reflects the method as actually run. This is a load-bearing concern because the headline 'up to +2.4%' is derived from the higher number. The recommended check is straightforward: reproduce Table I's config under the ablation protocol and report the mean. If the mean confirms the lower values, the headline should be adjusted; if the higher values are from a different setting, the paper must say so. This does not by itself invalidate the method, so I keep the reader's CONDITIONAL verdict.","tokens_in":12587,"tokens_out":10188,"duration_ms":115164,"concrete_test":"Ask the authors to (1) release the training logs for the Table I BEVFormer-tiny run and confirm whether it is the same configuration as Table III/IV's full framework; (2) rerun that exact configuration 5 times under the Table III protocol and report the mean±std. If the mean mAP is ≈0.2735–0.274 rather than 0.276, the '+2.4% mAP' headline should be revised to '+2.2%' and the large-model numbers to 0.422/0.525 instead of 0.424/0.528. If the original 0.276 is a selected best run, state that and use mean for claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of 'up to +2.4% mAP' is supported by Table I (BEVFormer-tiny: mAP 0.276, NDS 0.375) and repeated in Table V. However, the paper's own controlled ablation of the same configuration gives different results. In Table III, the 'Contrast Framework' row for ResNet-50 reports mAP 0.274 (mean 0.2735±0.0002 over 5 runs) and NDS 0.371 (mean 0.3709±0.0002). Table IV(a) last row, which includes all five components (Ins+Align+MLC+Pers+Scale), reports the same 0.274/0.371. The differences 0.002 mAP and 0.004 NDS are 10–20 standard deviations away from the reported variability, so they cannot be random noise. The same pattern appears for ResNet-101-DCN: Table II gives 0.424 mAP / 0.528 NDS, while Table III/IV(b) give 0.422 / 0.525. Thus the headline numbers do not match the repeated-run statistics that are supposed to represent the method. The paper does not explain whether Tables I/II are best-of-N runs, a different hyperparameter setting, or a different evaluation split. This is load-bearing because the quantitative magnitude of the contribution, and the claim of consistent gains, are calibrated to these larger numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces BEVCon, a training-time contrastive learning framework for BEV perception. Two auxiliary losses are added to a standard detection loss: an instance feature contrast loss computed on RoI-pooled BEV features from two augmented views, and a perspective regional contrast loss computed on multi-level image features. The method is evaluated on nuScenes 3D detection with BEVDet/BEVDet4D, BEVFormer-tiny/base, and Sparse4D, reporting consistent gains and an ablation study with repeated runs. The central claim is that these objectives, without extra data or labels, improve both the BEV encoder and the image backbone, giving up to +2.4 mAP over BEVFormer-tiny.","tokens_in":12993,"tokens_out":6131,"duration_ms":64752,"significance":"If the reported gains are reproducible, BEVCon would be a useful, architecture-agnostic training objective for multi-camera BEV detectors. The paper's strengths are its clear motivation, the breadth of detector families tested, and the fact that the main ablations are run multiple times with reported standard deviations. The code and models are promised open-source, which is valuable for verification. However, the central quantitative claim is weakened by an internal inconsistency between the headline tables and the controlled repeated-run tables, and the augmentation-alignment assumption underlying the contrastive pairs is underspecified. With those issues resolved, the contribution could be significant for the RA-L community.","major_comments":[{"comment":"The headline result is not internally reproducible. Table I reports BEVFormer-tiny+BEVCon as 0.375 NDS / 0.276 mAP, and Table II reports BEVFormer-base+BEVCon as 0.528 NDS / 0.424 mAP. However, the controlled ablation in Table III reports 'Contrast Framework' for ResNet-50 as 0.3709±0.0002 NDS / 0.2735±0.0002 mAP, and the full configuration in Table IV(a) (all five components) as 0.371 / 0.274. The differences (0.004 NDS, 0.002 mAP) are 10–20 times the reported standard errors, so they cannot be run-to-run noise. The same pattern holds for ResNet-101-DCN: Table II gives 0.528/0.424 vs 0.525±0.0004/0.422±0.0005 in Tables III/IV(b). The abstract's 'up to +2.4% mAP' is calibrated to the Table I value; the repeated-run numbers would support at most +2.2% mAP, and the NDS gain would be 1.7 points rather than 2.1. Please state which numbers are the official results, whether Tables I/II are bes","section":"§IV.D, Table III"},{"comment":"The ablation results do not support the sentence 'the combination of them could further improve the performance.' For ResNet-50, Instance Contrast alone gives NDS 0.375, Perspective Contrast alone gives 0.373, but Contrast Framework gives 0.371—lower than either individual module on NDS, despite a tiny mAP improvement (0.272→0.273→0.274). If NDS is the primary metric, the two modules are not complementary in this configuration. The same issue appears in Table IV(a), where the full model (0.371/0.274) does not dominate the 'Ins+Align+MLC' row (0.375/0.272) on NDS. Please report per-component significance and clarify whether the claimed synergy refers only to mAP.","section":"§IV.D, Table III"},{"comment":"The correctness of the method depends on exact BEV-space correspondence between the two augmented views: Eq. (1) treats fi and f'i as features of the same grid cell/instance. Section IV.B only asserts that 'the whole augmentation process ensures that each instance feature pair ... is located in the same position in BEV space', but does not specify how the 2D image augmentations (random resize, crop, rotation, distortion) are coupled with the BEV augmentations, nor how the 3D annotations are transformed for RoI pooling. A subtle misalignment—e.g., a per-camera resize that changes the projection—would make positive pairs mismatched and could actively harm training. Please provide the exact augmentation protocol (including alignment of 3D boxes with the BEV grid), and an empirical check such as the localization error of pooled features or a sensitivity study to small perturbations.","section":"§III.C-D, §IV.B"}],"minor_comments":[{"comment":"Baseline rows are reported without error bars, while the proposed method's repeated-run standard deviations are at the 10^-4 level. Please report the number of seeds and mean±std for the baselines as well, since the reported gains are small in several configurations (e.g., BEVDet4D +0.3 mAP).","section":"Tables I, II, III"},{"comment":"The checkmark formatting is difficult to parse; each row's component combination should be explicit (e.g., separate columns for Ins, Align, MLC, Pers, Scale with clear ✓/✗ entries). As printed, several rows are ambiguous.","section":"Table IV"},{"comment":"The sentence 'τ is set respectively to 0.2 and 0.05 for the two image backbones' should specify which backbone uses which value, and whether the same temperature is used for both contrast modules.","section":"§IV.B"},{"comment":"Define Nlayer explicitly and clarify whether the exponential scaling is applied to all transformer layers or only intermediate ones. No ablation or sensitivity study is provided for the hyperparameter ε.","section":"§III-C, Eq. (2)"},{"comment":"The 'ACO' dataset is not defined in the main text. Please give a reference and describe the pretraining setup (epochs, optimizer, views) so the comparison is reproducible.","section":"Table V"}],"recommendation":"major_revision","confidential_remarks":"The Table I/II vs Tables III/IV discrepancy is the key issue; if the larger numbers are not reproducible, the magnitude of the contribution drops considerably. I would ask the authors to either rerun the headline experiments with the same protocol as the ablation or clearly label the results as best-of-N. The augmentation-alignment concern also needs concrete implementation details before the method can be independently reproduced."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is simple and practical: add two contrastive losses—one on BEV instance features, one on perspective-view regional features—to standard detection training. It needs no new data or labels, and it shows positive mAP/NDS deltas across three detector families (BEVDet, BEVFormer, Sparse4D). The negative result on image-level contrast pre-training is a genuinely useful datapoint. The paper is clearly written, the code is linked, and the ablation coverage is decent. This is a legitimate empirical contribution.\n\nThe load-bearing problem is an internal reporting inconsistency. The abstract and Table I claim BEVFormer-tiny reaches 0.276 mAP / 0.375 NDS, which yields the '+2.4% mAP' headline. But the paper's own repeated-run statistics for the identical configuration (Tables III and IV) give 0.2735 ± 0.0002 mAP / 0.3709 ± 0.0002 NDS. That is 10–20 standard deviations away from the headline numbers. It cannot be noise. The paper never explains whether Tables I/II are best-of-N runs, a different checkpoint, or something else. This matters because the quantitative claim is calibrated to those larger numbers. If the real mean is 0.2735, the gain over the 0.252 baseline is +2.15 mAP, not +2.4—still positive, but the discrepancy itself needs to be addressed before the results can be trusted at face value.\n\nThe ablation story is also somewhat oversold. In Table IV, adding Perspective Contrast to Instance Contrast + Align + MLC drops NDS from 0.375 to 0.371, and the full framework's NDS stays below the instance-only setting. The paper says the combination 'further improves performance,' but that is only true for mAP, and only marginally. The perspective module's main documented benefit is stabilizing scale-aware pooling, which is a weaker claim.\n\nMinor reproducibility gaps: the authors assert that their augmentation preserves exact BEV-space alignment for positive pairs, but give no proof or detailed description. EMA momentum is also not reported. These are small issues, not fundamental flaws.\n\nOverall, the method is plausible and the gains are likely real, but the headline table conflict is a serious reporting flaw. The paper deserves a rigorous referee, but I would recommend insisting on a reconciliation of Tables I/II with Tables III/IV and a rewrite of the ablation conclusions to match what the numbers actually show.","headline":"Useful training-time contrastive recipe for BEV detectors, but the headline numbers don't match the paper's own repeated-run means, and the ablation story is shakier than the text claims.","tokens_in":13446,"tokens_out":2379,"would_cite":false,"duration_ms":28695,"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":"Adding two contrastive losses—one on BEV instance features, one on perspective image regions—improves 3D detection by up to 2.4 mAP on nuScenes without extra data or labels.","keywords":["bird's eye view perception","contrastive learning","3D object detection","autonomous driving","representation learning","nuScenes","multi-camera perception"],"falsifier":"Project every 3D ground-truth box center through the view transform into both augmented BEV grids $B$ and $B'$ and compute the fraction of instance pairs whose RoI-Align pooling regions cover the same grid cell; if that fraction is much less than 1, the alignment assumption fails and the reported gains cannot be attributed to the contrastive mechanism as described. Alternatively, train with independent random BEV rotations for the two views so positives are deliberately misaligned: a significant mAP drop relative to the aligned setting would confirm that alignment, not the contrastive objectiv","tokens_in":12497,"feed_emoji":"🚗","tokens_out":6188,"duration_ms":63924,"temperature":0.7,"pith_summary":"BEVCon claims that the bottleneck in camera-based Bird's Eye View perception is not just the encoder architecture or the detection head, but the quality of the learned feature representations themselves. The paper inserts two contrastive learning modules into existing BEV detectors: one pulls together BEV-pooled features of the same object from two augmented views while pushing apart different objects, and the other does the same for region-pooled features in the perspective camera images. Trained jointly with the detection loss on nuScenes, this raises mAP by up to 2.4 points (BEVFormer-tiny from 0.252 to 0.276) and improves accuracy across BEVDet, BEVDet4D, BEVFormer, and Sparse4D—families with different view transform mechanisms. The reason to care is that the gains come without new data, new labels, or architectural changes, suggesting representation-level supervision is a widely applicable complement to task-specific design. The paper also reports that image-level contrastive pre-training on driving data fails to help, which motivates the object-level, dense design.","feed_headline":"Two contrastive losses lift BEV detection up to 2.4 mAP","feed_subtitle":"Object-level contrast on BEV and camera features improves 3D detection across four architectures with no extra data.","key_machinery":"The load-bearing machinery is the pair of contrastive losses operating at object level. The instance feature contrast loss $\\mathcal{L}_{\\text{in}}$ uses RoI Align to pool features from BEV grids $B$ and $B'$ at the same ground-truth locations, forming positive pairs per instance and negatives across instances, with a SimCLR-style InfoNCE loss and cosine similarity; it is applied at multiple BEV encoder layers with an exponential scale $1/\\epsilon^{N_{\\text{layer}}-l}$. The perspective regional contrast loss $\\mathcal{L}_{\\text{pers}}$ pools multi-level image features with scale-aware pooling (shrinking boxes by $\\gamma=0.6$) and contrasts them across the two augmented views. An EMA-updated","core_discovery":"On the paper's own terms, the central discovery is that dense, annotation-guided contrastive learning—not generic image-level contrast—benefits BEV perception. The instance feature contrast module takes BEV features $B$ and $B'$ produced from two augmented views of the same frame, pools per-object features at ground-truth locations with RoI Align, and maximizes cosine similarity between matching instances while treating other instances as negatives. The perspective regional contrast module applies the same idea to multi-scale image features, using scale-aware pooling ($\\gamma=0.6$) to avoid fusing overlapping bounding boxes. Both losses are added to the detection loss with EMA-updated backbo","pith_inferences":["Editorial inference: the same instance-level contrastive losses could be extended to other BEV tasks—semantic segmentation, trajectory prediction, and planning—since the representation-level supervision is not tied to box detection; a natural test is attaching the modules to BEV segmentation or planning models.","The alignment assumption implies a portability constraint: for a BEV method whose view transform is not a fixed mapping (e.g., recurrent or sampling-based queries), the two augmented views may not yield identical BEV grid positions, so the positive-pair construction would need re-derivation rather than direct copying.","A cheap diagnostic extension would be to measure positive-pair feature similarity during training: if the contrastive loss is working as intended, same-instance pairs across augmented views should become substantially more similar than cross-instance pairs, and this gap should correlate with the mAP gain.","The gains reported on nuScenes may understate or overstate the value on datasets with different object density, resolution, or camera overlap; the framework's reliance on ground-truth boxes for pooling means label noise or missing annotations at train time could directly degrade the contrast signal."],"forward_implications":["If BEVCon's claim holds, any existing camera-based BEV detector can be improved without new data or labels by adding these two losses, as demonstrated on BEVDet, BEVDet4D, BEVFormer, and Sparse4D.","The consistent NDS and mAP gains across depth-based, dense-query, and sparse-query view transforms suggest the improvement is in shared representation quality rather than in any single architecture-specific component.","The reduction in mATE, mAOE, and mAVE errors implies contrastive supervision sharpens features relevant for localization, orientation, and velocity—properties that should transfer to downstream tracking and planning.","The failure of image-level contrastive pre-training/joint training indicates that object-level, annotation-anchored contrast is the operative ingredient, so future contrastive designs for driving perception should focus at instance/region level."],"supporting_citations":[{"why":"Supplies the SimCLR contrastive objective (cosine similarity, temperature-scaled InfoNCE) that both BEVCon modules adapt.","marker":"[10]"},{"why":"Baseline depth-unprojection BEV detector; BEVCon is applied on top of it and BEVDet4D from [42].","marker":"[5]"},{"why":"BEVFormer is the main dense-query baseline; its multi-layer BEV features receive the instance feature contrast loss.","marker":"[21]"},{"why":"Sparse4D serves as the sparse-query baseline, with instance features extracted from sparse query embeddings for contrastive learning.","marker":"[22]"},{"why":"RoI Align from Mask R-CNN is used to pool instance features from low-resolution BEV feature maps.","marker":"[41]"},{"why":"MoCo v2 is the image-level contrastive pretraining baseline that BEVCon is compared against, showing such methods underperform.","marker":"[44]"},{"why":"nuScenes is the dataset and benchmark on which all detection and ablation results are measured.","marker":"[15]"}],"fun_headline_variants":["BEVCon: two contrastive losses add 2.4 mAP","No extra data: contrastive learning boosts BEV mAP by 2.4","Contrastive pooling on BEV and camera adds 2.4 mAP","Object-level contrast on BEV features lifts detection 2.4 mAP"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the data-augmentation pipeline keeps each object in exactly the same BEV grid cell across the two augmented views; if cropping, rotation, or BEV-space augmentation breaks this correspondence, the positive pairs used by the instance contrast loss would be mismatched and the added loss would actively harm learning.","fun_headline_variants_meta":{"raw":{"variants":["BEVCon: two contrastive losses add 2.4 mAP","No extra data: contrastive learning boosts BEV mAP by 2.4","Contrastive pooling on BEV and camera adds 2.4 mAP","Object-level contrast on BEV features lifts detection 2.4 mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001455,"raw_usage":{"total_tokens":5679,"prompt_tokens":716,"completion_tokens":4963,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":4877}},"tokens_in":460,"tokens_out":4963,"duration_ms":36008,"temperature":1.0,"reasoning_tokens":4877,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:47:46.425096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Project every 3D ground-truth box center through the view transform into both augmented BEV grids $B$ and $B'$ and compute the fraction of instance pairs whose RoI-Align pooling regions cover the same grid cell; if that fraction is much less than 1, the alignment assumption fails and the reported gains cannot be attributed to the contrastive mechanism as described. Alternatively, train with independent random BEV rotations for the two views so positives are deliberately misaligned: a significant mAP drop relative to the aligned setting would confirm that alignment, not the contrastive objectiv","supporting_citations":[],"review_version":1}