{"id":"153acf7f-55c2-459d-b876-e83fe448d795","arxiv_id":"2505.07398","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"DepthFusion uses depth-based weighting in global and local cross-attention to fuse LiDAR and camera features, and it outperforms prior 3D detectors on nuScenes, KITTI, and nuScenes-C.","lead":"DepthFusion adds a depth signal to how self-driving perception models combine camera images with LiDAR point clouds, letting the model trust whichever sensor is more informative at each distance. The method reports state-of-the-art 3D detection scores on nuScenes and KITTI and is more robust to sensor corruption.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth-encoding ablation is not isolated: the w/o D variant drops depth multiplication together with the depth-gating pathway, so the attribution of the gain to depth-aware modulation is underdetermined.","rationale":"The reader identified the same central concern: the ablation evidence does not establish that the mechanism is depth-aware modulation rather than added capacity or gating. I agree with the CONDITIONAL verdict and with the emphasis on the missing control. My contribution is sharper: Eq. 3 shows the depth encoding is multiplied into the query, so the w/o D ablation removes the entire multiplicative interaction, not just the depth signal. This makes the attribution underdetermined in the very experiment that is cited as proof. I also note that the corruption-robustness explanation is unanchored: since no corruption-aware input exists, the claimed adaptive down-weighting under fog is inferred post-hoc. The accuracy claims, backbone-matched comparisons, and parameter-free design are solid and should be credited; the concern is about the interpretation, not about the numbers. A plausible alternative explanation is that any multiplicative interaction improves feature modulation, or that the extra nonlinearity in the attention query helps; even a fixed per-distance-bin scale would produce depth-dependent weights, which is much weaker than the claimed adaptive mechanism. The proposed constant-depth and permuted-depth controls would settle this directly. Therefore the verdict stays CONDITIONAL: accept the empirical contribution conditional on either releasing code or running the control ablation to substantiate the mechanism, and on toning down the unqualified 'first' claim. I found no internal inconsistency in the equations; the primary issue is evidential underdetermination of the central explanatory claim. My agreement with the reader is 'agree' because our diagnoses coincide; I did not need to change the verdict direction.","tokens_in":17733,"tokens_out":2178,"duration_ms":19826,"concrete_test":"Run an ablation where the depth matrix D is replaced by a constant (e.g., all elements set to the mean depth, or to 1.0) while keeping the same multiplicative query modulation and the same architecture. Also run a randomized-depth control: permute the depth values across BEV positions before encoding. If the randomized-depth variant retains the full gain of Table IX and the constant-depth variant also retains most of it, then performance does not depend on depth information and the 'depth-aware' attribution fails. If the query multiplication is removed exactly as in the w/o D ablation, that control does not separate the depth signal from the multiplicative operation, so the constant-depth and permuted-depth comparisons are the cleanest discriminators.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's explanatory claim is that the accuracy gain comes from depth-aware modulation of modality weights (Abstract, Sec. I; Eq. 3; Sec. III-B/C). The key evidence is Table IX, where removing D from DGF/DLF loses 1.5 NDS / 1.2 mAP. But D is not an optional input to the same attention; it enters as a multiplicative factor on the query: ((V_B^G + P) * D) in Eq. 3, and similarly in DLF for the query. Removing D therefore deletes the entire depth-related multiplicative interaction, not just a nonessential conditioning signal. The reported gain conflates at least three mechanisms: (1) genuine depth-dependent reweighting of image features, (2) the nonlinear feature transformation introduced by the elementwise product, and (3) a fixed multiplicative gating effect on feature scales regardless of depth variation. Table X comparing multiplication vs summation/concatenation does not separate these: multiplication may win simply because of its nonlinear scaling effect, not because it tracks depth. Fig. 8's attention-weight curves are post-hoc analysis; nothing in the training objective constrains weights to follow depth, so they do not settle the mechanism. The nuScenes-C robustness result is similarly explained post-hoc ('fog reduces image quality, so weight decreases'), but the model has no corruption-awareness input, so that explanation is not forced by any architectural property. What is solid: backbone-matched improvements on nuScenes/KITTI/nuScenes-C, consistent ablations, parameter-free depth encoding; the accuracy claim is well supported. The load-bearing weakness is narrow: the central explanatory claim of depth-adaptive fusion is not proven by the current ablation design, because the removed component changes the model's capacity and interaction structure, not just its depth information.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DepthFusion, a LiDAR-camera 3D object detection method built on the BEVFusion baseline. It introduces a parameter-free sine/cosine depth encoding of the BEV coordinates and uses it in two fusion modules: Depth-GFusion (DGF), which multiplies the query in a global cross-attention module by the depth encoding before querying image BEV features, and Depth-LFusion (DLF), which applies a similar depth-scaled attention to locally cropped raw voxel and image features. Experiments on nuScenes, KITTI, and nuScenes-C report consistent improvements over previous multimodal detectors when comparisons are grouped by 2D backbone, and ablations show that each module and the depth encoding contribute to the gains. The central claim is that depth-aware modulation of modality weights, rather than the added fusion capacity alone, is responsible for the improvements, especially for distant objects and under corruptions.","tokens_in":18162,"tokens_out":5472,"duration_ms":53560,"significance":"If the mechanism claim were fully supported, this would be a useful and practical contribution: a parameter-free depth signal that improves a strong baseline across multiple backbones, with clear benefits for far-range detection and robustness to corruptions. The paper's strengths include backbone-matched comparisons on nuScenes and KITTI, ablation of each module, evaluation on the nuScenes-C robustness benchmark, and a parameter-free depth encoding with no extra parameters. The main weakness is that the evidence for the depth-aware mechanism is partly underdetermined: the key ablation removes the entire multiplicative depth interaction, and the robustness explanation is based on post-hoc attention statistics rather than on a property enforced by the architecture. The absence of error bars and code release also limits confidence, though these are common in this line of work.","major_comments":[{"comment":"The w/o D ablation does not isolate depth-dependent modulation. In Eq. (3), the query is (V_B^G + P) · D, so removing D simultaneously removes (i) depth-dependent reweighting, (ii) the multiplicative feature transformation itself, and (iii) any fixed gating effect that D may induce on feature scales. The comparison in Table IX therefore conflates at least three mechanisms. Table X shows that multiplication beats summation and concatenation, but that result does not separate these mechanisms either, because multiplication may win simply through a nonlinear scaling effect that does not track depth. To support the paper's central explanatory claim, please add a control experiment in which D is replaced by a depth-independent gate (e.g., a scalar learned from the query features without depth, or a shuffled/random depth map) so that the accuracy gain can be attributed specifically to depth-consistent modulation.","section":"§III-B2, Eq. (3), Table IX"},{"comment":"The robustness explanation is post-hoc. The model has no corruption-awareness input, so the assertion in Sec. IV-J that 'the degradation of image feature quality leads to a reduction in the weight assigned to the image modality' is not forced by any architectural property; Fig. 8(a) shows attention statistics under different weather conditions but does not establish a causal link. Since the abstract and Sec. IV-D claim superior robustness as a contribution, this mechanism evidence is load-bearing. A stronger test would be to corrupt only the image branch or only the LiDAR branch during inference and measure whether the depth-dependent attention weights shift in the predicted direction, or to compare against a variant in which the depth gating is frozen when corrupted inputs are encountered.","section":"§IV-D, §IV-J, Fig. 8"},{"comment":"The far-distance comparison is internally inconsistent. The text says 'compared to BEVFusion [20]' and 'surpassing ObjectFusion by 13.4 pp', but Table V does not list ObjectFusion and the baseline shown is BEVFusion [21]. Please correct the citation and either add ObjectFusion's per-depth numbers or remove the unsupported margin. In addition, the protocol for depth-stratified mAP (how predictions are assigned to depth groups, and how matching is performed) should be specified so that the far-distance claim in Tables IV and V is reproducible.","section":"§IV-E, Table V"},{"comment":"No error bars or multi-seed results are reported, and no code is released. For a paper whose main quantitative claim is a few percentage points of NDS/mAP, a single run per configuration makes it difficult to judge whether the reported gaps are outside run-to-run noise. Reporting mean and standard deviation over at least three seeds for the key comparisons and ablations would substantially strengthen the empirical claim.","section":"General"}],"minor_comments":[{"comment":"The text reports 163.7 points per object within 0-10 meters, while the figure displays 163.9; please unify the numbers.","section":"Fig. 1 and Sec. I"},{"comment":"The sentence 'our DepthFusion-light our DepthFusion-light achieves 33.6% mAP' contains a duplicated phrase.","section":"§IV-G"},{"comment":"The phrase 'local feauture fusion' contains a typo; it should be 'local feature fusion'.","section":"§III-C"},{"comment":"The CMT row appears to have a formatting issue: the columns after FPS contain only three values, whereas other rows contain four; please verify and align the entries.","section":"Table I"},{"comment":"The axes and units of the attention-weight curves are not defined. Please specify what is averaged, over which dataset split, and how 'depth' is discretized.","section":"Fig. 8"},{"comment":"The row labeled 'BEVFusion + Ours' is unclear and does not correspond to a named model variant; please clarify which configuration this represents.","section":"§IV-H and Table VII"},{"comment":"The claim of being 'the first to observe' that modality roles vary with depth is stated without qualification; please discuss existing depth-aware fusion works (e.g., MSMDFusion's multi-depth seeds, GAFusion's guidance, 3D-DFM's dynamic filters) to clarify the precise novelty.","section":"Abstract and Sec. I"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a CV journal and the empirical architecture results are plausible. The main risk is that the central explanatory claim—that depth-aware modulation causes the gains—is underdetermined by the current ablations. I would encourage the editor to require the additional control experiments described in the major comments, rather than treating the mechanism as established. The 'first to observe' novelty claim also deserves more careful qualification against existing depth-guided fusion methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the empirical claims hold up, but the paper's broader story — that gains come from depth-aware modulation — is not actually proven by the ablations as designed. I'd send it to review, and I'd ask the authors to either add an experiment that isolates depth information or soften the mechanism language.\n\nWhat's new: two simple fusion modules (DGF, DLF) that inject a parameter-free sine-cosine depth encoding into global and local cross-attention. The design is clean and the gains are consistent: 1-3 NDS over strong backbones on nuScenes, competitive on KITTI, and notably better at >30m and under LiDAR dropout. The authors compare by 2D backbone, which is more careful than many papers. Ablations (Tables VIII-X) show each module and the depth encoding contribute, and the parameter-free encoding beats a learned one in their setting.\n\nWhere it's soft: the central explanatory claim. In Eq. 3, depth multiplies the query: (V_B^G + P)·D. Removing D doesn't remove depth information from the same attention; it removes the entire multiplicative interaction, so the 1.5 NDS drop in Table IX conflates (a) genuine depth-dependent reweighting, (b) a nonlinear feature transformation from the elementwise product, and (c) a fixed gating effect. Table X compares multiplication vs summation vs concatenation, but multiplication is also just a stronger nonlinear operation; it doesn't show the model tracks depth. Fig. 8 is post-hoc: nothing in training forces attention weights to follow depth. The nuScenes-C robustness story ('fog degrades images, so weight falls') is the same kind of after-the-fact narrative. None of this disproves the method — the accuracy numbers likely stand — but it does mean the paper's 'depth-aware' label is an interpretation, not a demonstrated mechanism.\n\nMinor: no code release, no error bars, and a small inconsistency between Fig. 1's printed value and the text (163.9 vs 163.7). The 'first to observe' claim is also stronger than warranted; depth-dependent utility is implicit in earlier adaptive fusion work.\n\nBottom line: worth a serious referee. The architecture is simple, reproducible in principle, and the gains look real. I'd ask for code, error bars, and either a cleaner depth-isolation experiment (e.g., fixed per-depth scaling vs. learned, or a depth-shuffled control) or a more careful claim. This is a good submission, not a great one.","headline":"Empirically solid, mechanistically underdetermined: the fusion gains look real, but the 'depth-aware' explanation is not actually isolated by the ablations.","tokens_in":18588,"tokens_out":2315,"would_cite":false,"duration_ms":21378,"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":"Depth-weighted fusion outperforms prior LiDAR-camera 3D detectors","keywords":["LiDAR-camera fusion","3D object detection","depth encoding","BEV features","cross-attention","nuScenes","KITTI","robustness"],"falsifier":"Train the full DepthFusion-base with the depth matrix replaced by a random but fixed permutation of depth values across BEV cells. If NDS stays near the reported 74.0 instead of dropping toward the no-depth ablation level of 71.8, the gains come from the gating machinery, not from depth information specifically.","tokens_in":17554,"feed_emoji":"🚗","tokens_out":3768,"duration_ms":34540,"temperature":0.7,"pith_summary":"DepthFusion claims that LiDAR-camera 3D object detection improves when the fusion of point cloud and image features is explicitly conditioned on depth. The paper shows statistically that LiDAR points dominate at close range while images become more informative at long range, and then builds a fusion strategy that lets the model weight each modality by depth. On nuScenes, KITTI, and the corrupted nuScenes-C benchmark, the reported accuracy and robustness exceed previous state-of-the-art methods. If the central claim holds, depth-aware weighting is a simple, parameter-free way to get more out of existing multimodal detectors.","feed_headline":"Depth-weighted fusion outperforms prior LiDAR-camera 3D detectors","feed_subtitle":"Modulating image vs LiDAR weight by depth improves far-range detection and corruption robustness.","key_machinery":"The depth encoding: a precomputed depth matrix storing the Euclidean distance of each BEV cell from the ego vehicle, converted into a parameter-free sine-cosine embedding following the positional-encoding recipe of the transformer. This embedding is multiplied element-wise into the point cloud BEV features that serve as the query in cross-attention, at both the global scale (Depth-GFusion) and the local instance scale (Depth-LFusion), thereby letting the attention weights modulate image-feature influence by depth. The ablation shows multiplication beats summation or concatenation for embedding the depth signal.","core_discovery":"The paper's central claim is that introducing depth encoding into both global BEV fusion and local instance-level fusion lets a LiDAR-camera detector adaptively emphasize whichever modality is more reliable at each distance, improving detection accuracy and robustness. Specifically, Depth-GFusion multiplies a sine-cosine depth encoding into the point cloud BEV query before cross-attending to image BEV features, so image influence grows with depth; Depth-LFusion applies the same depth-adaptive weighting to raw voxel and multi-view image instance features to recover detail lost in BEV projection. The authors report state-of-the-art NDS and mAP on nuScenes (74.9 NDS on validation for the largest variant), gains over prior methods on KITTI, and smaller performance drops under corruptions on nuScenes-C, with the largest accuracy improvements at distances beyond 30 meters.","pith_inferences":["The same depth-gating principle may transfer to other complementary sensor pairs, such as radar-camera or event-camera fusion, where measurement density also decays with range.","A testable extension would be applying only the depth-encoding multiplication to a single-modality detector, isolating whether the gain comes from modality weighting or from the added cross-attention capacity.","The robustness result hints that depth gating could be augmented with an explicit per-modality quality estimate, making the weighting respond not just to depth but to weather or sensor degradation.","If depth is truly the driving signal, then inverting the depth matrix (making near cells look far) should measurably degrade near-range detection; this prediction is a direct check of the paper's explanatory claim."],"forward_implications":["If correct, depth-adaptive fusion improves detection across all depth ranges, with the largest gains at far range (47.5 mAP beyond 30 meters versus 34.1 for the BEVFusion baseline).","The method degrades less under corruptions (a 4.63-point NDS drop on nuScenes-C versus 5.34 for the baseline), suggesting the depth weighting implicitly down-weights unreliable image features.","Because the depth encoding is parameter-free, the accuracy gains come with no additional learnable parameters for the depth signal, and the light variant runs at 13.8 FPS while beating the heavier baseline.","Combining global and local depth-aware fusion works better than either alone in ablations (73.3 NDS versus 72.4 and 72.7), indicating the two levels carry complementary information.","The success on small far-away objects such as traffic cones and barriers suggests depth-aware weighting directly addresses the sparse-point failure mode of LiDAR-only detection."],"supporting_citations":[{"why":"The BEVFusion baseline on which DepthFusion is built, providing the BEV fusion pipeline, image BEV projection, and the primary ablation comparison numbers.","marker":"[21]"},{"why":"The other BEVFusion version used as a main comparison point in depth-wise and far-distance evaluations.","marker":"[20]"},{"why":"The nuScenes dataset, the source of the statistical depth analysis and the main evaluation benchmark.","marker":"[1]"},{"why":"The nuScenes-C corruption benchmark that grounds the robustness claims.","marker":"[3]"},{"why":"TransFusion, a transformer-based fusion baseline compared in depth-grouped, far-distance, and LiDAR-dropout experiments.","marker":"[14]"},{"why":"IS-Fusion, the strongest same-configuration competitor that DepthFusion-large claims to surpass while running twice as fast.","marker":"[26]"},{"why":"SparseFusion, the source of the instance-level fusion design and the number of regressed 3D boxes used in Depth-LFusion.","marker":"[22]"},{"why":"The transformer paper that supplies the sine-cosine encoding recipe used to build the parameter-free depth encoding.","marker":"[28]"}],"fun_headline_variants":["Depth-aware fusion beats LiDAR-camera detectors","Depth-adaptive fusion boosts 3D detection and robustness","DepthFusion: depth-encoding adaptively blends LiDAR and camera features","Depth-modulated fusion improves far-range 3D detection","Depth-guided fusion improves LiDAR-camera 3D detection accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The extra accuracy from adding depth encoding is caused by the model genuinely modulating modality importance by depth, rather than by the extra cross-attention parameters or the multiplicative gating operation itself.","fun_headline_variants_meta":{"raw":{"variants":["Depth-aware fusion beats LiDAR-camera detectors","Depth-adaptive fusion boosts 3D detection and robustness","DepthFusion: depth-encoding adaptively blends LiDAR and camera features","Depth-modulated fusion improves far-range 3D detection","Depth-guided fusion improves LiDAR-camera 3D detection accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001077,"raw_usage":{"total_tokens":4503,"prompt_tokens":935,"completion_tokens":3568,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":3483}},"tokens_in":551,"tokens_out":3568,"duration_ms":25704,"temperature":1.0,"reasoning_tokens":3483,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:16:57.867830+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full DepthFusion-base with the depth matrix replaced by a random but fixed permutation of depth values across BEV cells. If NDS stays near the reported 74.0 instead of dropping toward the no-depth ablation level of 71.8, the gains come from the gating machinery, not from depth information specifically.","supporting_citations":[{"cited_title":"Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation","cited_arxiv_id":null,"evidence_quote":"The BEVFusion baseline on which DepthFusion is built, providing the BEV fusion pipeline, image BEV projection, and the primary ablation comparison numbers."},{"cited_title":"Bevfusion: A simple and robust lidar-camera fusion framework","cited_arxiv_id":null,"evidence_quote":"The other BEVFusion version used as a main comparison point in depth-wise and far-distance evaluations."},{"cited_title":"nuscenes: A multimodal dataset for autonomous driving","cited_arxiv_id":null,"evidence_quote":"The nuScenes dataset, the source of the statistical depth analysis and the main evaluation benchmark."},{"cited_title":"Benchmarking robustness of 3d object detection to common corruptions","cited_arxiv_id":null,"evidence_quote":"The nuScenes-C corruption benchmark that grounds the robustness claims."},{"cited_title":"Transfusion: Robust lidar-camera fusion for 3d object detection with transformers","cited_arxiv_id":null,"evidence_quote":"TransFusion, a transformer-based fusion baseline compared in depth-grouped, far-distance, and LiDAR-dropout experiments."},{"cited_title":"Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection","cited_arxiv_id":null,"evidence_quote":"IS-Fusion, the strongest same-configuration competitor that DepthFusion-large claims to surpass while running twice as fast."},{"cited_title":"Sparsefusion: Fusing multi-modal sparse representations for multi- sensor 3d object detection","cited_arxiv_id":null,"evidence_quote":"SparseFusion, the source of the instance-level fusion design and the number of regressed 3D boxes used in Depth-LFusion."}],"review_version":1}