{"id":"b0c60cf4-5334-422c-bc42-28a2f816b923","arxiv_id":"2505.13856","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A multi-modal camera-LiDAR framework with cross-attention fusion and three-level point-element interactions sets new state-of-the-art mAP on nuScenes and Argoverse2 HD map construction.","lead":"SuperMapNet is a neural network that fuses camera images and LiDAR point clouds to build vectorized high-definition maps for self-driving cars over a 120-meter range. It reports large accuracy gains over previous methods on two standard driving datasets, which matters because HD maps are central to autonomous vehicle perception and planning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claims rest on cross-protocol comparisons: prior results come from 60 m perception ranges and self-defined thresholds, so the reported mAP gaps may not reflect a fair head-to-head.","rationale":"I read the paper as a systems and empirical claim: the proposed coupling modules yield long-range, high-accuracy vectorized HD map construction, quantified by mAP on nuScenes and Argoverse2. The reader's conditional verdict centers on numerical comparability, and I agree this is the load-bearing assumption. The most consequential version of the concern is not only which method is second-best in one row, but that the evaluation ranges differ: all baseline numbers are cited from papers that evaluate at 60 m in Y while SuperMapNet evaluates at 120 m in Y. Because mAP is averaged over all ground-truth elements in the evaluated region, a model covering a larger region has a different denominator, so the published deltas are not differences of like quantities. The paper's own statements, 'existing methods used different AP thresholds for evaluation' and 'The results of the comparative methods are referenced in their paper,' are explicit limitations located in Section 4.1 and the Table captions. The hard-setting comparison adds a second issue: the strongest multi-modal baseline is absent under the hard thresholds, so the +14.9 mAP number is not against the strongest comparable method. I do not see an internal inconsistency in the network design, and the ablations are plausible, but the empirical headline is not fully established. The fix is to re-run at least the strongest baselines under SuperMapNet's own protocol. Until then, the appropriate verdict is conditional, which matches the reader's verdict; I therefore recommend UNCHANGED.","tokens_in":19263,"tokens_out":3978,"duration_ms":38476,"concrete_test":"Run the released code for SuperMapNet and at least HIMapNet (both C and C+L) on the nuScenes val split under three identical conditions: (1) the original 60 m Y range with SuperMapNet's hard/easy thresholds; (2) the full 120 m Y range with the same thresholds; and (3) per-element mAP reported at each range. If the 14.9/8.8 mAP margins shrink materially, or if HIMapNet C+L closes the gap when evaluated at 120 m, then the headline SOTA claim is not supported by the current tables.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is benchmark superiority: 66.5/86.6 mAP on nuScenes and 64.3/82.4 mAP on Argoverse2, with gains of 14.9/8.8 and 18.5/3.1 mAP over second-best methods. For this claim to hold, Table 1 and Table 2 must compare numbers produced under one evaluation protocol. The paper's own Section 4.1 violates this: SuperMapNet is evaluated at [-15,15] m in X and [-60,60] m in Y, and the text repeatedly describes 120 m in Y as twice the perception range of prior methods, yet all baseline entries are 'referenced in their paper' rather than re-evaluated at SuperMapNet's range. The paper also states that 'existing methods used different AP thresholds for evaluation' and then defines hard/easy threshold sets itself. Thus 120 m-range mAP values are being subtracted from 60 m-range mAP values, and the threshold protocol may differ as well. The hard-setting comparison is additionally weakened because the strongest multi-modal baseline, HIMapNet C+L, has no hard entry, so the +14.9 mAP gain is computed against camera-only HIMapNet. This does not disprove the architecture, but it means the headline margins are not established by the tables as printed; the claims need re-evaluation under a shared protocol before they can be taken at face value.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SuperMapNet, a camera-LiDAR framework for online vectorized HD map construction. It couples camera and LiDAR BEV features through a cross-attention synergy enhancement module and a flow-based disparity alignment module (SGC), and couples point queries with element queries through three interactions, Point2Point, Element2Element, and Point2Element (PEC). Experiments on nuScenes and Argoverse2 report 66.5/86.6 mAP and 64.3/82.4 mAP under the paper's hard/easy settings, which the authors interpret as 14.9/8.8 and 18.5/3.1 mAP gains over the second-best methods, at a perception range of 120 m in the Y-axis. The paper includes module-level ablations and threshold-sensitivity experiments.","tokens_in":19554,"tokens_out":4700,"duration_ms":43127,"significance":"If the reported numbers are taken at face value, the gains are large and the architecture is of genuine interest to the HD map construction community. The SGC module targets a real limitation of direct multi-modal concatenation, the PEC module explicitly models element-level relations that point-only methods ignore, the ablation study isolates the contribution of each module, and the code is publicly released. However, the central benchmark-superiority claim is not currently established because the comparisons mix evaluation protocols, and because the hard-setting gain is computed against a camera-only baseline. The architectural contribution remains plausible independently of the headline margins.","major_comments":[{"comment":"The headline SOTA margins are computed from numbers that are not demonstrably produced under a common protocol. Section 4.1 sets the perception range to [-15,15] m in X and [-60,60] m in Y and defines the hard/easy AP threshold sets {0.2,0.5,1.0} and {0.5,1.0,1.5}, while the table captions state that all comparative results are 'referenced in their paper.' The paper itself notes in Section 1 and Section 4.2 that other methods operate at roughly 60 m in Y, and in Section 4.1 that existing methods used different AP thresholds. If the cited baseline numbers were obtained at a 60 m range or under different AP thresholds, subtracting them from SuperMapNet's 120 m-range mAP is not a valid head-to-head comparison. Because the abstract and Section 4.2 state the gains as the central claim, the authors should either re-evaluate the baselines under the same range and threshold protocol or explicitly restrict the claim to a cross-protocol comparison and discuss the direction and magnitude of the resulting bias.","section":"§4.1, Tables 1 and 2"},{"comment":"In the hard setting, the 'second-best' used to compute the +14.9 mAP gain is camera-only HIMapNet (51.6 mAP), because the multi-modal HIMapNet C+L row has no hard-setting entry ('-'). The gain statement does not disclose that the comparison is against a camera-only method rather than the strongest comparable multi-modal baseline. This weakens the hard-setting SOTA claim as printed. The authors should report a hard-setting result for HIMapNet C+L under the same protocol, or revise the claim and its wording to state explicitly which baseline the gain is measured against.","section":"§4.2, Table 1"},{"comment":"The 'long-range' claim is not supported by any range-stratified evaluation. The paper reports aggregate mAP over a 120 m Y-range and repeatedly contrasts this with the 60 m range of prior methods, but Table 1 and Table 2 contain no per-range breakdown and no same-protocol baseline at 60 m. Without such data, the reader cannot tell how much of the reported margin comes from the SGC module's long-range fusion and how much from the different evaluation range. A complementary experiment evaluating SuperMapNet and at least one baseline at both 60 m and 120 m would make the long-range contribution testable.","section":"§4.2, Tables 1 and 2"},{"comment":"The threshold ablation shows that the combination of SGC and PEC is substantially better than either module alone only at the strict 0.2 m threshold, while at 1.0 m and 1.5 m it matches PEC-only performance. This is an honest and useful observation, but the text in Section 5.2 that says the advantage 'basically does not exist, and achieves the same accuracy as only using PEC module, as the SGC module' is garbled: Table 4 shows SGC-only is lower at those thresholds, so the final phrase should be corrected to 'not as the SGC module' or reworded.","section":"§5.2, Table 4"}],"minor_comments":[{"comment":"The disparity alignment equations contain an indexing/typographical error: Eq. (11) defines weight_w' = max(0, 1 - |h + Δh - h|), which reduces to |Δh| and cannot be the intended weight for the w-coordinate. The source and target coordinate indices (h, w, h', w', and any vertical-flow variable) should be defined consistently before Eqs. (10)-(12).","section":"§3.2, Eqs. (10)-(12)"},{"comment":"In Table 2, the ADMap row is cited as 'ADMap (Wilson et al., 2021)', but ADMap is the method of Hu et al. (2024); the citation is incorrect and should be fixed.","section":"Table 2"},{"comment":"The statement that SuperMapNet 'maintains real-time capability' at 5.0 FPS is not self-evident for an online HD map construction system. If 5 FPS is considered real-time in this application context, the criterion should be stated explicitly; otherwise the claim should be softened to a latency report.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The central issue is empirical comparability. The paper's own statement in Section 4.1 that existing methods used different AP thresholds, combined with the table captions saying baseline results are 'referenced in their paper', indicates that the reported SOTA margins are cross-protocol differences rather than controlled comparisons. I would recommend requiring either same-protocol re-evaluations of the baselines or a substantially qualified statement of the gains, before the claims in the abstract can be taken at face value. The architectural contribution and ablations are solid enough to justify a major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about SuperMapNet. The architecture work is real: the semantic-geometric coupling (cross-attention plus flow alignment) and the three-level point-element coupling are clearly motivated, and the ablations show each module adds non-trivial mAP over a reasonable multi-modal baseline. The code is public. But the headline SOTA numbers—14.9/18.5 mAP gains—are not trustworthy as printed. The authors evaluate at a 120 m Y-range while every prior result they cite was produced at a 60 m Y-range, and they define their own threshold sets after noting that 'existing methods used different AP thresholds.' A longer range means a different ground-truth set, so the mAP gaps in Tables 1 and 2 are cross-protocol comparisons, not head-to-head wins. The hard-setting 'second-best' is also a camera-only method (HIMapNet) on both datasets, because the strongest multi-modal baselines (HIMapNet C+L) don't report hard numbers; the easy-setting gains against multi-modal methods are smaller (8.8/3.1) and more believable. Internally, the paper is solid: SGC adds 4.9/3.2 and PEC 7.7/6.7 mAP over a concatenation baseline, and the strict-threshold ablations are sensible. There's also a small citation error: Table 2 lists ADMap as 'Wilson et al., 2021,' but the reference correctly attributes it to Hu et al., 2024. Read this if you care about vectorized HD map construction or camera-LiDAR BEV fusion; the module designs and ablations are worth a look. My recommendation for an editor: send it to review—the method deserves scrutiny and the code is available—but make the authors re-run or properly re-calibrate baselines under a single evaluation protocol, or soften the SOTA claim to what the evidence supports.","headline":"Sensible architecture and honest ablations, but the headline SOTA margins rest on cross-protocol comparisons and do not hold up as printed.","tokens_in":20098,"tokens_out":3419,"would_cite":false,"duration_ms":31392,"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":"SuperMapNet couples camera semantics with LiDAR geometry and point-level with element-level modeling to construct vectorized HD maps over a 120 m range, reporting state-of-the-art mAP on nuScenes and Argoverse2.","keywords":["vectorized HD map","multi-modal fusion","bird's-eye-view","LiDAR point clouds","cross-attention","point-element coupling","long-range perception","autonomous driving"],"falsifier":"Run the strongest prior camera-plus-LiDAR method under SuperMapNet's own protocol, with a 120 m Y-axis range, 0.15 m voxelization, hard thresholds {0.2, 0.5, 1.0} m and easy thresholds {0.5, 1.0, 1.5} m, on the same splits, and compute mAP; if hard/easy mAP meets or exceeds 66.5/86.6 on nuScenes or 64.3/82.4 on Argoverse2, the SOTA claim collapses. Alternatively, replace SGC and PEC in SuperMapNet with plain concatenation and a point-only decoder under the same 120 m range; equal accuracy would falsify the mechanism claim.","tokens_in":19068,"feed_emoji":"🗺️","tokens_out":9593,"duration_ms":80514,"temperature":0.7,"pith_summary":"SuperMapNet's claim is that vectorized HD map construction can be pushed to a 120 m perception range and to higher accuracy by treating fusion and decoding as structured coupling problems rather than concatenation. On the fusion side, it couples camera semantics with LiDAR geometry using cross-attention synergy enhancement plus flow-based disparity alignment. On the decoding side, it couples point-level and element-level modeling through three interactions: Point2Point, Element2Element, and Point2Element. The reported results on nuScenes and Argoverse2, new best mAP under both hard and easy thresholds, would matter because autonomous driving needs maps that are both long-range and accurate at the centimeter-to-decimeter level.","feed_headline":"SuperMapNet doubles map range to 120 m and beats prior SOTA","feed_subtitle":"Two coupling modules lift HD map accuracy on nuScenes and Argoverse2 at twice the usual range.","key_machinery":"The load-bearing machinery is the pair of coupling modules. The SGC module has two sub-modules: a cross-attention synergy enhancement that treats camera BEV features and LiDAR BEV features as query/key/value pairs in both directions, so each modality's features absorb complementary information from the other; and a flow-based disparity alignment that predicts a per-cell displacement $(\\Delta h, \\Delta w)$ and bilinearly resamples the camera BEV features onto LiDAR coordinates before the final concatenation. The PEC module operates on ordered keypoint sets with learnable point queries and element queries, and uses three attention levels, Point2Point, Element2Element, and Point2Element, where Point2Element updates point descriptors with element descriptors and vice versa via cross-attention with position embedding. These modules implement the paper's two premises: fusion must manage cross-modal synergy and sensor disparity, and decoding must use both point-level and element-level information.","core_discovery":"The central claim is that two specific design choices explain high-accuracy long-range vectorized HD mapping. First, the semantic-geometric coupling (SGC) module fuses camera and LiDAR BEV features not by direct concatenation but by cross-attention, letting each modality fill the other's gaps, followed by a learned flow-based alignment that registers camera features onto the more accurate LiDAR coordinates before concatenation. Second, the point-element coupling (PEC) module decodes map elements with three levels of interaction: Point2Point for intra-element geometry, Element2Element for inter-element semantic constraints, and Point2Element for passing global element context back to points. With both modules, SuperMapNet reports 66.5/86.6 mAP on nuScenes and 64.3/82.4 mAP on Argoverse2 under its hard/easy AP thresholds, corresponding to claimed gains of 14.9/8.8 and 18.5/3.1 mAP over the second-best compared methods, over a 120 m Y-axis range that is twice the prior 60 m.","pith_inferences":["A natural extension the paper does not test is whether the flow-based disparity alignment is robust to deliberate miscalibration; injecting synthetic translation or rotation errors between camera and LiDAR during evaluation would show whether the module actually corrects sensor disparity or merely learns dataset-specific offsets.","Because the hard-setting second-best in Table 1 is camera-only, the hard-setting margin may shrink if the strongest multi-modal baseline publishes a hard-setting number; the easy setting is where a camera-plus-LiDAR comparison exists.","The degradation of SGC's benefit at thresholds above 0.5 m, if replicated, implies sensor alignment is a precision enabler rather than a general accuracy booster; systems targeting meter-level maps could skip disparity alignment and save compute.","Point2Point's weak effect on pedestrian crossings, which are modeled with only two keypoints, suggests a testable modification: allocate more keypoints or an alternative primitive for short elements."],"forward_implications":["If the 120 m operating point is adopted by later systems, downstream planning modules can receive map information twice as far ahead, changing when lane-change and braking decisions can be made.","The claimed gains imply that direct concatenation of camera and LiDAR BEV features leaves measurable accuracy unused, specifically about 4.9/3.2 mAP on nuScenes validation for SGC alone.","The PEC ablation suggests element-level context is worth more than the fusion module alone on this task, with +7.7/+6.7 mAP, and is especially valuable for elements modeled by many keypoints.","At the strict 0.2 m threshold, the combined SGC+PEC gain is +11.1 mAP, so the modules matter most where centimeter-level accuracy is demanded."],"supporting_citations":[{"why":"Provides the nuScenes dataset with multi-sensor data and map annotations used for training and evaluation.","marker":"(Caesar et al., 2020)"},{"why":"Provides the Argoverse2 dataset with LiDAR, cameras, and 3D vectorized maps.","marker":"(Wilson et al., 2021)"},{"why":"Supplies the ordered keypoint-set representation and dynamic keypoint matching that SuperMapNet adopts for map element modeling.","marker":"(Ding et al., 2023)"},{"why":"HIMapNet is the main multi-modal baseline and the hybrid point-element query design that PEC extends; also the source of the Argoverse2 train/val/test split protocol.","marker":"(Zhou et al., 2024)"},{"why":"PointPillars converts LiDAR point clouds into BEV features through dynamic voxelization.","marker":"(Lang et al., 2019)"},{"why":"Swin Transformer serves as the shared backbone that encodes multi-view camera image features.","marker":"(Liu et al., 2021)"},{"why":"Deformable attention lifts perspective camera features into the unified BEV space.","marker":"(Zhu et al., 2021)"},{"why":"MapTRv2 supplies the hybrid-query baseline that combines element and point queries; PEC positions itself against this design by adding three explicit interaction levels.","marker":"(Liao et al., 2024)"}],"fun_headline_variants":["SuperMapNet uses cross-attention to build 120m HD maps","Two coupling modules double HD map range and boost accuracy","SuperMapNet: cross-modal fusion for 120m vector HD maps","Semantic-geometric coupling extends HD map range to 120m"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported state-of-the-art gains assume that the earlier results quoted for comparison used the same evaluation rules, same map coverage windows, same data splits, and same matching thresholds, so that the accuracy gaps compare like with like.","fun_headline_variants_meta":{"raw":{"variants":["SuperMapNet uses cross-attention to build 120m HD maps","Two coupling modules double HD map range and boost accuracy","SuperMapNet: cross-modal fusion for 120m vector HD maps","Semantic-geometric coupling extends HD map range to 120m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000375,"raw_usage":{"total_tokens":2071,"prompt_tokens":1086,"completion_tokens":985,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":702,"completion_tokens_details":{"reasoning_tokens":911}},"tokens_in":702,"tokens_out":985,"duration_ms":7723,"temperature":1.0,"reasoning_tokens":911,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:08:27.401162+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the strongest prior camera-plus-LiDAR method under SuperMapNet's own protocol, with a 120 m Y-axis range, 0.15 m voxelization, hard thresholds {0.2, 0.5, 1.0} m and easy thresholds {0.5, 1.0, 1.5} m, on the same splits, and compute mAP; if hard/easy mAP meets or exceeds 66.5/86.6 on nuScenes or 64.3/82.4 on Argoverse2, the SOTA claim collapses. Alternatively, replace SGC and PEC in SuperMapNet with plain concatenation and a point-only decoder under the same 120 m range; equal accuracy would falsify the mechanism claim.","supporting_citations":[],"review_version":1}