{"id":"147dad3a-5063-48b4-9719-3cc5320b7950","arxiv_id":"2502.04377","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A plug-in attention-and-gating fusion method improves multi-modal HD map and BEV map construction by a few points on nuScenes and Argoverse2.","lead":"MapFusion adds two fusion modules to a camera-LiDAR map construction pipeline: a self-attention step that lets the two sensor modalities interact in bird's-eye view, and a gating fusion step that picks the more informative features. On nuScenes, it reports 3.6% higher HD-map mAP and 6.2% higher BEV-segmentation mIoU than the baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline provenance undermines the claimed SOTA gains: Table 1's MapTR LiDAR/C&L numbers are not from the camera-only MapTR paper, so the +3.6 HD-map improvement is measured against an unvalidated control.","rationale":"The central claim is empirical: CIT+DDF improve multi-modal BEV fusion enough to beat state-of-the-art methods by 3.6 mAP on HD map construction and 6.2 mIoU on BEV map segmentation on nuScenes. For that claim to be meaningful, the control baselines must be real, well-tuned, and produced under equal settings. The reader identified this baseline-equivalence assumption as the weakest; my stress-test agrees and finds textual support for an even sharper problem. Table 1 claims all numbers are 'taken from MapTR [25]', but the original MapTR is camera-only, so the LiDAR and C&L rows are not from that paper, and Section 2.1 incorrectly groups MapTR among camera-LiDAR fusion methods. These rows look like the authors' own unpublicized variants, possibly from the ICRA predecessor MBFusion. If the LiDAR branch in that control is under-tuned, the +3.6 gain is not evidence for the modules. The BEV segmentation comparison uses official BEVFusion numbers and is better anchored, but it still lacks error bars and relies on author re-implementations on Argoverse2. The paper does have supporting evidence: monotonic ablations, plug-and-play gains on HDMapNet and VectorMapNet, and t-SNE visualizations showing feature alignment. These do not, however, validate the absolute improvement over a questionable control. The proposed check, independently reproducing the two baseline conditions, would settle whether the gains are attributable. My recommended verdict remains conditional as the reader's: the method is plausible and internally coherent, but the evidence is not yet sufficient without code release and baseline verification.","tokens_in":21654,"tokens_out":12270,"duration_ms":110729,"concrete_test":"Reproduce the two control conditions that anchor the headline gains: (i) MapTR C&L with ResNet-50 + SECOND on nuScenes for 24 epochs using the official MapTR codebase extended with a LiDAR branch exactly as described in Section 4.3; verify mAP is approximately 62.5 within a tolerance of 0.5. (ii) BEVFusion C&L with Swin-T + VoxelNet on nuScenes for 20 epochs using the official BEVFusion code; verify mIoU is approximately 62.7 within 0.5. If baseline (i) reproduces above 63.5 or below 61.5, the +3.6 mAP claim is not attributable to the proposed modules; if (ii) deviates by more than 1 point, the +6.2 mIoU claim is similarly compromised. The authors should also release both baseline and full model configs to enable this check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline improvement of +3.6 mAP on HD map construction is measured against a control condition whose provenance is questionable. Table 1 states that all numbers are 'taken from MapTR [25]', but MapTR (ICLR 2023) is a camera-only method; the LiDAR-only and camera+LiDAR rows (55.6 and 62.5 mAP) cannot come from that paper. Indeed, Section 2.1 itself misclassifies MapTR as a camera-LiDAR fusion method. These baselines appear to be the authors' own unpublicized variants, likely inherited from their ICRA predecessor MBFusion, not an independently published state of the art. If the LiDAR branch in that control was integrated naively (e.g., simple concatenation, which the paper criticizes), the +3.6 gain may reflect a weak control rather than the CIT/DDF modules. Similar risk affects Table 9, where the authors' HDMapNet re-implementation (28.2 mAP) falls below the published 31.0, suggesting their re-implementations are systematically under-tuned. The BEV segmentation claim (+6.2 mIoU) is comparatively safer because the BEVFusion C&L baseline is official, but it still relies on single-run results without variance. Until the MapTR C&L baseline is independently reproduced or the code is released, the central claim of superiority on HD map construction is not adequately supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MapFusion, a camera-LiDAR BEV feature fusion method for map construction, consisting of two modules: a Cross-modal Interaction Transform (CIT) that applies multi-head self-attention to the concatenated camera and LiDAR BEV tokens, and a Dual Dynamic Fusion (DDF) module that uses gated weighting and squeeze-excitation to fuse the two feature streams. The authors evaluate MapFusion on HD map construction and BEV map segmentation on nuScenes and Argoverse2, reporting absolute improvements of +3.6 mAP, +4.1 mAP, +6.2 mIoU, and +6.5 mIoU over their baselines, along with ablation studies, compatibility experiments, and visualizations. The paper is an extension of the authors' ICRA 2024 conference paper and adds the BEV segmentation task, additional ablations, and further analysis.","tokens_in":21963,"tokens_out":6233,"duration_ms":57148,"significance":"If the reported results are reproducible, the proposed modules constitute a simple and potentially useful plug-and-play contribution to multi-modal BEV fusion for map-related tasks. The compatibility study in Table 9 and the t-SNE and feature-map visualizations provide supporting evidence that the modules can be inserted into existing pipelines. However, the evaluation infrastructure currently prevents acceptance of the headline claims: the provenance of the key HD-map baseline is questionable, there is an internal backbone inconsistency in the BEV segmentation ablation, and all results are single runs without variance information. The method itself is not mathematically derived, but the empirical claims are the main contribution, so these evaluation issues are load-bearing.","major_comments":[{"comment":"The caption of Table 1 states that the comparison numbers are 'taken from MapTR [25]', but MapTR (ICLR 2023) is a camera-only method; the LiDAR-only (55.6 mAP) and camera-LiDAR (62.5 mAP) rows cannot come from that publication. Section 2.1 also lists MapTR among 'camera-LiDAR fusion methods', which is a misclassification. Because the headline +3.6 mAP improvement on HD map construction is measured against this unvalidated control, the central claim is not adequately supported. Please provide the exact origin of the LiDAR and C&L MapTR numbers, specify how those variants were implemented (including the fusion operation and training schedule), and release code or model weights so the control can be reproduced.","section":"Table 1 and Section 2.1"},{"comment":"The BEV segmentation ablation in Table 6 labels the backbone as 'ResNet-50 & VoxelNet', while Table 3 and Section 4.3 state that the BEV segmentation experiments use Swin-T and VoxelNet. Moreover, the baseline row in Table 6 reports the same mIoU (62.7) as the Swin-T baseline in Table 3, which cannot be true if the backbone differs. This internal inconsistency affects the interpretation of the +6.2 mIoU improvement and the component ablation. Please clarify the actual backbone and report a single consistent set of results.","section":"Table 6 vs. Table 3 and Section 4.3"},{"comment":"All results are reported as single runs with no variance information. Several conclusions hinge on small differences, e.g., Add Fusion vs. Dynamic Fusion on HD map construction (64.4 vs. 64.3 mAP in Table 7) and Add Fusion vs. Baseline on BEV segmentation (+0.2 mIoU in Table 8). Without multi-seed statistics, these differences and even some of the claimed component gains may not be statistically meaningful. Please report mean and standard deviation over at least three seeds for the main results and ablations.","section":"Tables 5-9"},{"comment":"The Argoverse2 results (Tables 2 and 4) and the compatibility results (Table 9) rely on author re-implementations marked with †, with no code release or detailed configuration files. The HDMapNet† re-implementation in Table 9 (28.2 mAP) is below the published HDMapNet C&L result (31.0 mAP in Table 1), which suggests the re-implemented controls may be systematically under-tuned. Please provide the exact training configs, logs, and code for the re-implemented baselines, and compare with any available published numbers.","section":"Tables 2, 4, and 9"}],"minor_comments":[{"comment":"The sentence 'Training losses include classification loss, point2point loss, and edge direction loss. with weights of 2.0, 5.0, and 0.005' has a punctuation error, and the learning-rate description 'starting for 1e-4 and performing a single cycle with target ratios 10, 1e-4 and a step of 0.4' is unclear. Please rewrite for precision.","section":"Section 4.3"},{"comment":"The correlation matrix in Eq. 5 has misaligned row and column labels and is difficult to read. Please reformat the matrix so that the intra-modality and inter-modality blocks are clearly labeled.","section":"Equation 5"},{"comment":"The M2BEV row contains '%' placeholders for several classes. Please either fill in the actual values or remove the row to avoid presenting incomplete data.","section":"Table 3"},{"comment":"The description of the AvgPool operations is confusing: Eq. 6 says AvgPool outputs a vector of length C, while the remarks say the AvgPool in Eq. 8 outputs a W x H map, and Eq. 8 then applies a linear transform W to that map. Please clarify the exact tensor shapes and the role of the linear transform.","section":"Section 3.4 remarks"},{"comment":"The conclusion repeatedly describes CIT as a 'cross-attention mechanism', but Section 3.3 defines it as self-attention over the concatenated BEV tokens. Please align the terminology throughout the paper.","section":"Section 5 and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper is an extension of the authors' ICRA 2024 paper, and the incremental novelty is mainly the additional BEV segmentation task and expanded ablations. The baseline-provenance problem in Table 1 appears to be inherited from the conference version and must be resolved before the claims can be assessed. The backbone inconsistency between Table 6 and Table 3 suggests that two different experimental setups may have been conflated; the authors should be asked to rerun and report a single consistent set of experiments. I would also recommend requiring code release or detailed configuration files, given the number of re-implemented baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the fusion recipe is sensible and the ablations are internally consistent, but the headline HD-map gain is measured against a MapTR camera+LiDAR baseline that the cited MapTR paper never published. That provenance problem is load-bearing.\n\nWhat's actually new: not the components—CIT is multi-head self-attention over concatenated BEV tokens with a learned positional embedding, and DDF is a squeeze-and-excitation gate on the summed features plus a conv fusion. Both are known building blocks. The new part is putting them together for BEV map construction and evaluating the combination as a plug-in on two tasks and two datasets. The ablation tables (5-8) consistently show each module adds something, and the compatibility table (9) shows boosts on HDMapNet, VectorMapNet, and MapTR. That's useful evidence for a practical technique.\n\nWhere it gets soft: Table 1 says all baselines are 'taken from MapTR [25]', but MapTR is a camera-only method. The LiDAR-only and camera+LiDAR rows (55.6 and 62.5 mAP) cannot come from that paper. Section 2.1 even lists MapTR under camera-LiDAR fusion. These numbers are likely inherited from the authors' ICRA predecessor, and they're unvalidated. If the LiDAR branch in that control was a weak concatenation, the +3.6 mAP improvement reflects the control, not necessarily the proposed modules. The stress-test note is right about this. Similarly, the Argoverse2 re-implementations (marked †) are unverified, and the HDMapNet re-implementation in Table 9 (28.2 mAP) is below the published 31.0, which hints their re-implementations may be under-tuned.\n\nThe BEV segmentation claim is safer: the BEVFusion camera+LiDAR baseline is official, and the +6.2 mIoU is a real comparison. Still, there are no error bars or multi-seed runs, so single-run differences of a few points are not conclusive.\n\nThere's also a minor internal inconsistency: Table 6 labels the backbone ResNet-50, while Table 3 uses Swin-T for the same baseline. Probably a typo, but it needs fixing.\n\nBottom line: the paper is a solid engineering contribution with an honest ablation story, but the main HD-map comparison is built on a shaky control. I'd send it to peer review—the flaws are fixable—but I'd tell reviewers to demand the code, the exact provenance of the MapTR LiDAR/C&L numbers, and variance estimates. The method itself deserves a shot in a revised form.","headline":"Useful fusion recipe with consistent ablations, but the HD-map headline gain rests on a MapTR baseline the cited paper never published.","tokens_in":22497,"tokens_out":3126,"would_cite":true,"duration_ms":27839,"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":"The paper proposes MapFusion, a plug-and-play camera-LiDAR BEV fusion scheme whose Cross-modal Interaction Transform and Dual Dynamic Fusion modules improve HD map construction mAP by 3.6% and BEV map segmentation mIoU by 6.2% on nuScenes.","keywords":["BEV feature fusion","multi-modal map construction","camera-LiDAR fusion","cross-modal interaction","self-attention","HD map construction","BEV map segmentation","autonomous driving"],"falsifier":"Retrain MapTR and BEVFusion under MapFusion's exact settings, including the same epochs, batch size, learning rate, and data pre-processing, and compare mAP and mIoU; if the gap drops below roughly one point or disappears, the improvements are not caused by the CIT and DDF modules.","tokens_in":21440,"feed_emoji":"🗺️","tokens_out":12628,"duration_ms":100761,"temperature":0.7,"pith_summary":"MapFusion is a camera-LiDAR feature fusion method for the two main autonomous-driving map tasks: vectorized high-definition map construction and bird's-eye-view (BEV) map segmentation. The paper's argument is that projecting camera and LiDAR features into the same BEV grid does not remove their semantic misalignment, and that the element-wise addition, weighted averaging, or concatenation used by previous fusion methods leaves that misalignment uncorrected. MapFusion therefore inserts two plug-in modules between the modality encoders and the map decoder: a Cross-modal Interaction Transform (CIT) that runs multi-head self-attention over the concatenated camera and LiDAR BEV tokens, and a Dual Dynamic Fusion (DDF) module that gates the two feature maps with complementary attention weights before merging them. Reported results on nuScenes show absolute gains of 3.6% mAP for HD map construction and 6.2% mIoU for BEV map segmentation over the corresponding baselines, with similar gains on Argoverse2, and the modules also improve HDMapNet, VectorMapNet, and MapTR when added on top. The paper concludes that explicitly aligning the two modalities inside the fusion stage is worth more than the naive fusion operations that current pipelines use.","feed_headline":"Bird's-eye map fusion gains 3.6-6.2 points with cross-modal attention","feed_subtitle":"Two plug-in modules align camera and LiDAR features in shared space, boosting HD map and BEV segmentation accuracy.","key_machinery":"The load-bearing machinery is the CIT-DDF pair inserted between the modality encoders and the map decoder. CIT flattens the camera and LiDAR BEV features into tokens $\\mathbf{T}^{BEV}_{Camera}$ and $\\mathbf{T}^{BEV}_{LiDAR}$, concatenates them with a learnable positional embedding, and applies multi-head self-attention; the correlation matrix $\\alpha$ splits into intra-modality and inter-modality blocks, so the self-attention simultaneously aligns each modality with itself and with the other. DDF takes the CIT outputs, generates a channel attention vector $\\mathbf{w}$ from their sum via global average pooling and linear layers with a sigmoid, multiplies the camera features by $\\mathbf{w}$ and the LiDAR features by $1-\\mathbf{w}$, concatenates them, and applies a $3\\times3$ convolution followed by an adaptive channel-spatial gate. This mechanism carries the paper's claim: CIT removes the semantic misalignment that naive fusion leaves in place, and DDF prevents information loss by softly selecting which modality to trust at each channel.","core_discovery":"On the paper's own terms, the central discovery is that the residual error in multi-modal map construction after BEV fusion is not a resolution or capacity problem but a correspondence problem: camera and LiDAR features at the same BEV grid location can represent different semantic content because the view transformation is depth-imperfect and the sensor modalities differ. The paper's solution, the Cross-modal Interaction Transform (CIT), treats the two feature maps as one token sequence and learns a correlation matrix $\\alpha = \\mathrm{softmax}(\\mathbf{Q}\\mathbf{K}^{\\top}/\\sqrt{D_k})$ whose four natural blocks are camera-to-camera, LiDAR-to-LiDAR, camera-to-LiDAR, and LiDAR-to-camera attention. This lets each spatial position be re-weighted by evidence from the other modality before fusion, which the paper demonstrates with t-SNE visualizations showing the two feature populations moving closer together after CIT. The Dual Dynamic Fusion (DDF) module then computes a channel-wise gate from the summed features and applies it to the two inputs in complementary form ($\\mathbf{w}$ and $1-\\mathbf{w}$), so the network can suppress the weaker modality where the stronger one is informative. Together these modules form MapFusion, and the paper reports that this interaction-then-selection design outperforms its baselines by the margins above on two tasks and two datasets.","pith_inferences":["Because CIT is defined purely on BEV tokens, the same alignment mechanism could transfer to other multi-modal BEV tasks such as 3D object detection or occupancy prediction; the paper states such an extension as future work without testing it.","A per-condition evaluation could test whether CIT's benefit grows as camera depth estimates become less reliable, since the paper's misalignment story predicts the largest gains where the camera-to-BEV projection is weakest.","DDF's complementary gate could be refined from channel-level weights to per-location routing, letting the network decide which modality owns each BEV cell; that is a natural next step the paper leaves implicit."],"forward_implications":["On nuScenes HD map construction, MapFusion reaches 66.1 mAP, beating MapTR's 62.5, with the largest single-class gain on pedestrian crossings (61.6 vs 55.9).","On nuScenes BEV map segmentation, MapFusion reaches 68.9 mIoU, beating BEVFusion's 62.7, and improves every class except car park.","The improvements transfer to Argoverse2: +4.1 mAP for HD map construction and +6.5 mIoU for BEV map segmentation over re-implemented baselines.","Adding MapFusion to re-implemented HDMapNet and VectorMapNet and to MapTR yields +7.6, +5.5, and +3.6 mAP respectively on nuScenes.","Ablations attribute the larger part of the segmentation gain to CIT (+5.6 mIoU alone) and a smaller part to DDF (+1.4), with the full model combining both."],"supporting_citations":[{"why":"Supplies the HD map construction baseline, the map head losses, and the nuScenes comparison numbers MapFusion is measured against.","marker":"[25]"},{"why":"Supplies the BEV map segmentation baseline, the segmentation head, and the comparison numbers for the fusion experiments.","marker":"[31]"},{"why":"Provides one of the three HD map methods used for comparison and compatibility tests, and the evaluation framework with Chamfer-distance-based average precision.","marker":"[21]"},{"why":"Provides the second HD map baseline and is used in the compatibility experiments showing MapFusion adds 5.5 mAP on top.","marker":"[28]"},{"why":"Provides a representative prior cross-modal fusion method that the paper contrasts as neglecting modality interaction.","marker":"[2]"},{"why":"Supplies the multi-head self-attention formulation that CIT adapts for cross-modal BEV alignment.","marker":"[43]"},{"why":"Serves as the default 2D-to-BEV transformation in the BEV map segmentation experiments.","marker":"[36]"},{"why":"Inspires the DDF attention-gating design, which the paper modifies into a dual complementary gate.","marker":"[19]"},{"why":"Provides the primary benchmark dataset on which the headline HD map and BEV segmentation numbers are reported.","marker":"[3]"},{"why":"Provides the second benchmark dataset used for the transfer results on both tasks.","marker":"[47]"}],"fun_headline_variants":["MapFusion: Cross-modal attention lifts HD map and BEV accuracy by 6.2%","BEV fusion with cross-modal interaction improves HD map segmentation by 6.2%","Two plug-in modules align camera-LiDAR features, boosting map accuracy","Cross-modal interaction and adaptive fusion enhance bird's-eye map construction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume that MapTR, BEVFusion, and the re-implemented Argoverse2 baselines were trained with the same backbone, schedule, augmentation, and evaluation protocol as MapFusion, so the 3.6 to 6.5 point differences come from the proposed modules rather than from training disparities.","fun_headline_variants_meta":{"raw":{"variants":["MapFusion: Cross-modal attention lifts HD map and BEV accuracy by 6.2%","BEV fusion with cross-modal interaction improves HD map segmentation by 6.2%","Two plug-in modules align camera-LiDAR features, boosting map accuracy","Cross-modal interaction and adaptive fusion enhance bird's-eye map construction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000311,"raw_usage":{"total_tokens":1855,"prompt_tokens":1110,"completion_tokens":745,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":726,"completion_tokens_details":{"reasoning_tokens":659}},"tokens_in":726,"tokens_out":745,"duration_ms":6923,"temperature":1.0,"reasoning_tokens":659,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:05:00.295604+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain MapTR and BEVFusion under MapFusion's exact settings, including the same epochs, batch size, learning rate, and data pre-processing, and compare mAP and mIoU; if the gap drops below roughly one point or disappears, the improvements are not caused by the CIT and DDF modules.","supporting_citations":[{"cited_title":"Maptr:Structuredmodelingandlearningforonlinevectorized HD map construction, in: International Conference on Learning Representations","cited_arxiv_id":null,"evidence_quote":"Supplies the HD map construction baseline, the map head losses, and the nuScenes comparison numbers MapFusion is measured against."},{"cited_title":"Bevfusion:Multi-taskmulti-sensorfusionwithunifiedbird’s-eye view representation, in: ICRA, IEEE","cited_arxiv_id":null,"evidence_quote":"Supplies the BEV map segmentation baseline, the segmentation head, and the comparison numbers for the fusion experiments."},{"cited_title":"Hdmapnet: An online hd map construction and evaluation framework, in: IEEE International Conference on Robotics and Automation, pp","cited_arxiv_id":null,"evidence_quote":"Provides one of the three HD map methods used for comparison and compatibility tests, and the evaluation framework with Chamfer-distance-based average precision."},{"cited_title":"Vectormapnet: End-to-end vectorized hd map learning, in: International Conference on Machine Learning, pp","cited_arxiv_id":null,"evidence_quote":"Provides the second HD map baseline and is used in the compatibility experiments showing MapFusion adds 5.5 mAP on top."},{"cited_title":"X-align: Cross-modal cross-view alignment forbird’s-eye-viewsegmentation,in:ProceedingsoftheIEEE/CVFWinterConferenceonApplicationsofComputerVision,pp.3287–3297","cited_arxiv_id":null,"evidence_quote":"Provides a representative prior cross-modal fusion method that the paper contrasts as neglecting modality interaction."},{"cited_title":"Attention is all you need, in: Advances in Neural Information Processing Systems, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-head self-attention formulation that CIT adapts for cross-modal BEV alignment."},{"cited_title":"Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d, in: European Conference on Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"Serves as the default 2D-to-BEV transformation in the BEV map segmentation experiments."},{"cited_title":"Squeeze-and-excitation networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp","cited_arxiv_id":null,"evidence_quote":"Inspires the DDF attention-gating design, which the paper modifies into a dual complementary gate."},{"cited_title":"11618–11628","cited_arxiv_id":null,"evidence_quote":"Provides the primary benchmark dataset on which the headline HD map and BEV segmentation numbers are reported."},{"cited_title":"Argoverse 2: Next generation datasets for self-driving perception and forecasting, in: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks","cited_arxiv_id":null,"evidence_quote":"Provides the second benchmark dataset used for the transfer results on both tasks."}],"review_version":1}