{"id":"bdbf1eca-0d87-4985-b3ba-312c2e4b6b2c","arxiv_id":"2502.01856","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"ReliFusion fuses LiDAR and camera BEV features with confidence-weighted mutual cross-attention, reporting state-of-the-art nuScenes accuracy and larger robustness gains under simulated LiDAR and camera failures.","lead":"ReliFusion is a LiDAR-camera 3D object detector that uses learned confidence scores to re-weight how much each sensor contributes, aiming to keep working when LiDAR or camera data is degraded or missing. The paper reports better robustness than BEVFusion when LiDAR is severely limited, but the evaluation protocol is not fully disclosed and the gains on clean data are marginal.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline protocol in Tables 2–3 is under-specified and inconsistent with Table 1, so the claimed robustness margin may stem from training/temporal settings rather than the reliability mechanism.","rationale":"The external comparison is the load-bearing evidence for the paper's central claim. The internal inconsistencies in clean numbers are factual and verifiable, making this a concrete flaw rather than a matter of consensus. The paper gives no unified evaluation protocol in Section 4.2, and the baselines' drop in clean performance relative to their published results strongly suggests a different training regime or a different number of temporal frames. The ablation study (Table 4) shows the modules help under the paper's own protocol, but that does not establish superiority over prior work. Because the central claim of SOTA robustness is unsupported without a fair comparison, the REJECT verdict is appropriate. If the authors release code and a unified protocol, the verdict could change, but as submitted the protocol gap is load-bearing.","tokens_in":11320,"tokens_out":4621,"duration_ms":44956,"concrete_test":"Run a controlled comparison on nuScenes val: take the official BEVFusion and TransFusion checkpoints and evaluate them under ReliFusion's exact protocol—same multi-frame camera inputs (T frames), same image size 448x800, same LiDAR corruption masks (FOV truncation and 50% box drop), and same detection head evaluation. Verify that their clean mAP matches published values; if not, adjust settings until it does. If the mAP gap at (-pi/2,pi/2) remains approximately 6 points after this alignment, the robustness claim survives; if it shrinks to near the clean-condition gap, the claim is an artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that ReliFusion outperforms prior fusion methods under LiDAR corruption—rests entirely on Tables 2 and 3 (Section 4.4). But the baselines' 'Clean' rows are inconsistent with Table 1: CenterPoint drops from 60.3 to 56.8 mAP, TransFusion from 68.9 to 66.9, BEVFusion from 69.2 to 67.9. This indicates the compared baselines are not the published checkpoints or are run under different settings. Section 4.2 specifies only ReliFusion's implementation; it never states what backbones, temporal frames, image resolution, or corruption simulation were used for CenterPoint, TransFusion, BEVFusion, PointAugmenting, or MVX-Net. In particular, ReliFusion's STFA (Sec. 3.2) aggregates camera BEV features across T frames, while the baselines are not described as temporal. If the baselines are single-frame, then under zero-LiDAR (FOV (-0,0)) the 24.6 vs 12.4 mAP gap would be explained by temporal camera information alone, not by the Reliability module. Consequently, the claimed robustness margin (52.4 vs 46.4 at (-pi/2,pi/2)) is confounded by protocol, and the SOTA claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes ReliFusion, a BEV-space LiDAR-camera fusion detector for nuScenes. Three components are introduced: STFA (spatial and temporal attention over camera BEV features), a Reliability module built on cross-modality contrastive learning that outputs per-modality confidence scores, and CW-MCA that weights cross-attention fusion by these scores. Experiments on the nuScenes test set report 70.6 mAP / 73.2 NDS, and robustness experiments under limited LiDAR FOV and object-region point drops report higher mAP than BEVFusion and other baselines. Ablation tables decompose the contributions of STFA, CW-MCA, and the Reliability module. The authors conclude that ReliFusion adaptively balances modalities and is robust to sensor malfunctions.","tokens_in":11618,"tokens_out":8068,"duration_ms":75391,"significance":"If the empirical claims were properly controlled, ReliFusion would address a practically important failure mode: maintaining 3D detection performance under partial or complete LiDAR loss. The architecture is plausible and the ablation design is useful. However, the paper currently does not establish its central claim. The robustness comparison in Tables 2 and 3 is invalidated by unspecified and inconsistent baseline protocols; the supervision target for the confidence loss is undefined; and the contrastive training data for the reliability module is not described. These are load-bearing gaps rather than presentation issues. The paper would benefit from releasing code and checkpoints to enable verification.","major_comments":[{"comment":"The baseline protocol is under-specified and the clean numbers are inconsistent with Table 1. For example, CenterPoint drops from 60.3 to 56.8 mAP, TransFusion from 68.9 to 66.9, BEVFusion from 69.2 to 67.9, and PointAugmenting from 66.8 to 46.9 between Table 1 and Tables 2/3. Section 4.2 details only ReliFusion's implementation (ConvMixer, VoxelNet, 448x800 images), and no statement is made about which backbones, temporal windows, or corruption simulation were used for the baselines. Since the STFA module (Section 3.2) aggregates camera features over T frames while the baselines are not described as temporal, the limited-FOV margins (52.4 vs 46.4 at (-pi/2,pi/2), 24.6 vs 12.4 with zero LiDAR) could be due to temporal camera information or training settings rather than to the reliability-driven fusion. The central robustness claim is therefore not established.","section":"Section 4.4, Tables 2 and 3"},{"comment":"The confidence loss L_conf has no defined supervision target. The scores C_LiDAR and C_Camera are defined in Eq. (12) as sigmoid outputs of learned projections, and L_conf is described only as 'penalizes inaccurate confidence predictions,' but no ground-truth confidence label is defined anywhere in the paper. The first-stage 'regression loss' for pre-training the Confidence Module is likewise unspecified. Without a target, the reliability scores may simply be hidden parameters that improve the detection objective, and the claim that they 'quantify the dependability of each modality' is circular rather than validated. The authors should define the target (for example, the degree of injected corruption) and show that the scores track corruption severity.","section":"Sections 3.3 and 3.5, Eq. (16)"},{"comment":"The contrastive learning setup is incompletely specified. Positive pairs are said to be 'unaltered data' and negative pairs 'corrupted data,' but Section 4.2 does not describe any corruption augmentation used during training, nor does it state whether the robustness-test corruptions are seen in pretraining. If the model is trained only on clean data, the negatives in Eq. (11) are other camera embeddings in the batch and the loss does not learn a reliable/corrupted distinction. A concrete training-time corruption protocol is required for the reliability module to be meaningful.","section":"Section 3.3, Eq. (11)"},{"comment":"The same experimental condition is reported with different numbers. For zero LiDAR FOV, Table 2 gives ReliFusion 24.6/28.7 (mAP/NDS), while Table 4 gives 24.6/39.7; for (-pi/3,pi/3), Table 2 gives 44.9/54.6 and Table 4 gives 44.9/54.8. The NDS differences are unexplained and make the ablation comparisons internally inconsistent. Please reconcile the reported numbers.","section":"Tables 2 and 4"}],"minor_comments":[{"comment":"The camera-failure protocol is ambiguous: the text says the front (F) image was 'entirely removed or preserved,' but it does not state what happens to the other five cameras in each condition; the 'Preserve F' condition should be defined explicitly.","section":"Section 4.4, Table 3"},{"comment":"There is a typographical error: '47.8.0/59.4' for MVX-Net under Object Failure should presumably be '47.8/59.4'.","section":"Table 3"},{"comment":"The notation '(-0, 0)' is confusing; it should be written as the empty set or as '(0,0)' with a clear definition of the zero-FOV protocol.","section":"Section 4.4, Tables 2 and 4"},{"comment":"The heading 'Object Detection Failures' is misleading; the experiment drops points inside object bounding boxes, so a heading such as 'LiDAR Object-Region Drop' would be more accurate.","section":"Table 2"},{"comment":"The paper credits MMDetection3D to reference [32], which is the MMDetection paper; either use the correct framework reference or clarify the relationship.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to mix evaluation splits across tables without disclosure: the clean baseline numbers in Tables 2 and 3 differ substantially from Table 1, which is consistent with a test/val mismatch. I recommend that the editor require the authors to state the evaluation split for every table and to provide the exact baseline evaluation protocol (backbones, temporal frames, and corruption code). The missing definition of L_conf is the most serious methodological gap and should be addressed before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI read ReliFusion. The core idea is genuinely new: a contrastively-trained reliability module that produces confidence scores to weight mutual cross-attention between LiDAR and camera BEV features, with a temporal aggregation module on the camera side. The ablations are internally consistent and each component adds something, especially under LiDAR degradation. That is real work.\n\nThe problem is that the central robustness comparison is not trustworthy as reported. The baseline clean numbers in Tables 2 and 3 differ from Table 1 for the same methods (TransFusion 66.9 vs 68.9, BEVFusion 67.9 vs 69.2, CenterPoint 56.8 vs 60.3). The paper never states whether the baselines use the same backbones, the same number of temporal frames, or the same corruption simulation. ReliFusion's STFA module aggregates camera BEV features over T frames; the baselines are not described as temporal. So the headline \"24.6 mAP with no LiDAR vs 12.4 for BEVFusion\" likely reflects STFA's temporal advantage at least as much as the reliability mechanism. The ablation table shows STFA alone gets 17.2 mAP at zero FOV, which already beats BEVFusion's 12.4. So the claim that the reliability module drives the robustness margin is not established.\n\nSecond, the confidence scores are never validated as measures of dependability. L_conf is mentioned but its supervision target is never defined. The scores are trained inside the detection network to improve fusion; they may be useful scalars, but the paper's claim that they \"quantify reliability\" is circular without comparison to ground-truth corruption levels.\n\nWhat is good: the ablation tables 4-6 are clear and the design choices are sensible. The clean-set results are competitive, though the gain over CMT is 0.2 mAP/NDS. The camera robustness experiments follow the RoboBEV setup, which is a plus.\n\nThis paper deserves a serious referee, but in current form it should not be accepted. The authors need to release code, fix the baseline protocol, define L_conf's target, and validate the confidence scores against actual corruption severity. If they do that, the reliability-module idea could be a solid contribution.\n\nI'd bring it to a reading group as a case study in how evaluation protocol can undermine a plausible idea. My call: send to review, expect major revision.","headline":"Useful robustness study undermined by an inconsistent baseline protocol and an unvalidated confidence module; the core idea is worth a serious look but the current evidence does not support the SOTA robustness claim.","tokens_in":74,"tokens_out":2708,"would_cite":false,"duration_ms":48896,"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":"ReliFusion fuses LiDAR and camera with per-sensor confidence weights so 3D detection survives sensor failures.","keywords":["3D object detection","LiDAR-camera fusion","sensor reliability","bird's-eye view","cross-modality contrastive learning","confidence-weighted cross-attention","temporal feature aggregation","nuScenes"],"falsifier":"Re-run the Table 2 and Table 3 comparisons with BEVFusion and TransFusion using ReliFusion's backbone, training schedule, temporal frames, and corruption pipeline; if either baseline matches or exceeds ReliFusion's mAP under limited LiDAR FOV, zero LiDAR, or camera occlusion, the reliability-driven advantage is not confirmed.","tokens_in":11166,"feed_emoji":"🚗","tokens_out":11015,"duration_ms":90506,"temperature":0.7,"pith_summary":"ReliFusion is a LiDAR-camera fusion framework for 3D object detection in autonomous driving. It is built around the idea that a fusion model should know when one of its sensors is unreliable: the network estimates a confidence score for the LiDAR stream and for the camera stream, then scales each modality's contribution to a mutual cross-attention step by that score. The authors train these confidence scores with contrastive learning that pulls clean LiDAR-camera pairs together and pushes corrupted pairs apart. On the nuScenes benchmark they report accuracy competitive with or above current fusion methods on clean data, and clearly higher mAP than BEVFusion when the LiDAR field of view is narrowed or removed entirely, as well as under camera failure and occlusion. The practical target is graceful degradation: a blinded or disconnected sensor should reduce detection quality smoothly instead of collapsing the detector.","feed_headline":"ReliFusion doubles detection mAP under total LiDAR loss","feed_subtitle":"Per-sensor confidence weights keep 3D object detection working when one sensor drops out.","key_machinery":"The central mechanism is the confidence-weighted mutual cross-attention (CW-MCA) fusion rule, fed by the contrastively trained Reliability module. In CW-MCA the camera features query the LiDAR features and the LiDAR features query the camera features; each attention output is multiplied by a scalar confidence score, $C_{\\text{Camera}}$ or $C_{\\text{LiDAR}}$, produced by a sigmoid over a learned projection of that modality's contrastive embedding. The two confidence-scaled streams are added to form the fused BEV feature that goes into the detection head. Before fusion, the Spatio-Temporal Feature Aggregation (STFA) module applies self-attention across the six camera views and across time steps, so the image stream has a stable multi-frame representation that can carry detection when LiDAR is reduced.","core_discovery":"The paper's central claim is that the brittleness of existing LiDAR-camera fusion under sensor malfunction comes from fixed fusion weighting, and that replacing it with confidence-weighted fusion restores robustness. ReliFusion's Reliability module embeds LiDAR and camera BEV features in a shared space, uses cross-modality contrastive learning to make clean pairs agree and corrupted pairs disagree, and maps each embedding to a sigmoid confidence score. The Confidence-Weighted Mutual Cross-Attention module then computes attention in both directions between the modalities and multiplies each direction's output by the corresponding confidence, so degraded input is down-weighted before features are summed. With a LiDAR field of view truncated to nothing, the authors report 24.6 mAP against 12.4 for BEVFusion; with the front camera missing or object pixels occluded, the drop from clean performance is a few mAP points. The conclusion they draw is that the same framework can serve both clean and degraded sensing without a separate failure-detection network.","pith_inferences":["A testable extension the paper does not run: attach the reliability module and CW-MCA to other BEV fusion backbones; if their degraded-LiDAR mAP rises by the same margin, the weighting mechanism rather than the specific backbone is the active ingredient.","The confidence scores could be exported to downstream driving modules as a per-frame trust signal, letting a planner slow down or switch strategies when the camera becomes the only reliable modality.","Because the contrastive head was trained on clean-versus-corrupted pairs from this task's corruption types, it is an open question whether unseen corruptions such as rain, snow, or lens dirt also separate from the clean cluster; that is directly measurable with the paper's pipeline."],"forward_implications":["Narrowing the LiDAR field of view degrades ReliFusion's detection gradually, with reported mAP of 52.4 at $(-\\pi/2, \\pi/2)$, 44.9 at $(-\\pi/3, \\pi/3)$, and 24.6 when no LiDAR remains.","Dropping the front camera or occluding half the object pixels costs only a few mAP points, since the LiDAR geometry still anchors the fused representation.","The reliability scores are trained end-to-end with detection, contrastive, temporal, and confidence losses, so no separate failure classifier is needed at runtime.","Under the authors' ablations, each added component -- STFA, CW-MCA, and reliability scoring -- raises limited-FOV and object-failure performance, with the largest relative gains appearing when LiDAR is most degraded."],"supporting_citations":[{"why":"Provides the nuScenes benchmark and sensor suite on which all detection and robustness numbers are measured.","marker":"[10]"},{"why":"The BEVFusion baseline whose limited-FOV and zero-LiDAR results are the main robustness comparison.","marker":"[7]"},{"why":"TransFusion supplies the detection head reused in ReliFusion and is a strong fusion baseline in Tables 1 through 3.","marker":"[24]"},{"why":"Lift-Splat-Shoot is the view transformation that projects multi-camera image features into BEV space for fusion.","marker":"[19]"},{"why":"Defines the camera-failure and object-occlusion setups reused in the camera robustness experiments.","marker":"[9]"},{"why":"CenterPoint sets the LiDAR voxelization and center-based detection paradigm adopted by the LiDAR stream and serves as a degradation baseline.","marker":"[30]"},{"why":"VoxelNet is the 3D backbone used to convert point clouds into BEV LiDAR features.","marker":"[12]"},{"why":"ConvMixer is the 2D backbone used to extract multi-view image features.","marker":"[34]"}],"fun_headline_variants":["ReliFusion doubles mAP under total LiDAR loss","Sensor down? ReliFusion keeps 3D detection on point","Confidence-weighting future-proofs LiDAR-camera fusion against failures","ReliFusion achieves double mAP when LiDAR vanishes entirely"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The robustness comparison assumes that the baseline methods were run under exactly the same protocol as ReliFusion, meaning the same backbones, the same number of temporal frames, and the same corruption simulation.","fun_headline_variants_meta":{"raw":{"variants":["ReliFusion doubles mAP under total LiDAR loss","Sensor down? ReliFusion keeps 3D detection on point","Confidence-weighting future-proofs LiDAR-camera fusion against failures","ReliFusion achieves double mAP when LiDAR vanishes entirely"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000763,"raw_usage":{"total_tokens":3382,"prompt_tokens":938,"completion_tokens":2444,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":2368}},"tokens_in":554,"tokens_out":2444,"duration_ms":17384,"temperature":1.0,"reasoning_tokens":2368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:12:30.110155+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Table 2 and Table 3 comparisons with BEVFusion and TransFusion using ReliFusion's backbone, training schedule, temporal frames, and corruption pipeline; if either baseline matches or exceeds ReliFusion's mAP under limited LiDAR FOV, zero LiDAR, or camera occlusion, the reliability-driven advantage is not confirmed.","supporting_citations":[{"cited_title":"nuscenes: A multimodal dataset for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Provides the nuScenes benchmark and sensor suite on which all detection and robustness numbers are measured."},{"cited_title":"Bevfusion: A simple and robust lidar-camera fusion framework,","cited_arxiv_id":null,"evidence_quote":"The BEVFusion baseline whose limited-FOV and zero-LiDAR results are the main robustness comparison."},{"cited_title":"Transfusion: Robust lidar-camera fusion for 3d object detection with transformers,","cited_arxiv_id":null,"evidence_quote":"TransFusion supplies the detection head reused in ReliFusion and is a strong fusion baseline in Tables 1 through 3."},{"cited_title":"Lift, splat, shoot: Encoding im- ages from arbitrary camera rigs by implicitly unproject- ing to 3d,","cited_arxiv_id":null,"evidence_quote":"Lift-Splat-Shoot is the view transformation that projects multi-camera image features into BEV space for fusion."},{"cited_title":"Benchmark- ing the robustness of lidar-camera fusion for 3d object detection,","cited_arxiv_id":null,"evidence_quote":"Defines the camera-failure and object-occlusion setups reused in the camera robustness experiments."},{"cited_title":"Center-based 3d object detection and tracking,","cited_arxiv_id":null,"evidence_quote":"CenterPoint sets the LiDAR voxelization and center-based detection paradigm adopted by the LiDAR stream and serves as a degradation baseline."},{"cited_title":"V oxelnet: End-to-end learning for point cloud based 3d object detection,","cited_arxiv_id":null,"evidence_quote":"VoxelNet is the 3D backbone used to convert point clouds into BEV LiDAR features."}],"review_version":1}