{"id":"03c5f53d-4a85-4d99-830f-37178284f6dc","arxiv_id":"2505.08808","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"SparseMeXt, a sparse-query model for HD map construction, reaches 68.9% mAP on nuScenes at over 20 FPS, marginally outperforming the dense MapTRv2 baseline.","lead":"This paper introduces SparseMeXt, a sparse representation network for real-time HD map construction from cameras that reaches 68.9% mAP on the nuScenes benchmark, narrowly beating the dense MapTRv2 baseline. It combines three modifications: a single-level image encoder, a sparse-to-dense segmentation auxiliary task, and physically inspired query denoising, showing sparse methods can compete with dense BEV approaches.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sparse-over-dense claim rests on an unmatched backbone comparison: SparseMeXt-Large is ResNet-101 while MapTRv2-Base is ResNet-50, and at matched ResNet-50 the paper's own numbers put the dense method ahead.","rationale":"The paper is a systematic attempt to close the sparse/dense gap in online HD map construction, and several individual ingredients (SiMo neck, nuImages pretraining, auxiliary segmentation, PPDN) are plausible and individually show positive deltas. I credit the authors for including a long-range experiment and a centerline task. However, the strongest claim, 'first sparse paradigm to surpass dense methods,' depends on a single 0.2-point margin in the least controlled row of Table VII: SparseMeXt-Large uses ResNet-101, while MapTRv2-Base uses ResNet-50. The paper's own matched-backbone comparison goes the other way, and the text contradicts the table about the MapTRv2-Base value. This is a correctness-risk issue, not a disagreement with the community's preference for dense BEV methods. The reader identified the fragility of single-run ablation deltas; my concern is related but distinct, focusing on the headline comparison itself. I would keep the conditional verdict: the method may well be competitive, but the central superiority claim needs either a dense ResNet-101 baseline, a tempering of the claim, or a matched-backbone demonstration before it can be accepted. No code is released, so reproducing the exact 68.9 number is also not currently possible.","tokens_in":11096,"tokens_out":9000,"duration_ms":83132,"concrete_test":"Run MapTRv2-Base with a ResNet-101 backbone under the same training recipe as SparseMeXt-Large (or with equivalent or greater capacity and training budget) on the nuScenes val split and compare mAP with 68.9. If the dense ResNet-101 model reaches or exceeds 68.9 mAP, the claim that sparse representations surpass dense methods is not supported. In the same audit, resolve the Section IV-B/Table VII discrepancy by deciding whether the MapTRv2-Base mAP is 63.0 or 68.7 and rerunning all comparisons consistently.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SparseMeXt is the first sparse paradigm to surpass dense methods in online HD map construction. In Table VII the only sparse-over-dense win is SparseMeXt-Large (ResNet-101, 68.9 mAP) against MapTRv2-Base (ResNet-50, 68.7 mAP). This is an unmatched comparison. The paper's matched-backbone entries show SparseMeXt-Base (ResNet-50, 65.2 mAP) behind MapTRv2-Base (ResNet-50, 68.7 mAP), i.e., the dense method leads by 3.5 points when the backbone is held fixed. No dense ResNet-101 baseline is reported, so the headline result is consistent with a capacity advantage rather than a sparse-paradigm advantage. The same paragraph in Section IV-B states MapTRv2-Base as 63.0%, while Table VII lists it as 68.7%; either the table or the text is wrong, and the magnitude of the claimed 0.2-point advantage is squarely within single-run uncertainty. Table X's stepwise gains of 0.07, 0.6, 0.5 and 0.9 points also have no variance or seed information, but the decisive problem is that the strongest claim is not supported by a controlled comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SparseMeXt, an online HD map construction architecture that retains a sparse, query-based paradigm instead of a dense BEV feature grid. It introduces three main components: an architecture optimized for map features (SiMo image encoder neck, adjusted temporal/non-temporal decoder stage ratio, and a decoupled deformable feature aggregation decoder), a query-centric sparse-dense auxiliary segmentation loss, and a physical-prior query denoising module (PPDN) with rotation, location, scale, and curvature noise. Experiments on nuScenes report SparseMeXt-Tiny/B/L at 55.5%, 65.2%, and 68.9% mAP with 32.9/25.4/20.2 FPS, respectively, plus long-range and centerline detection extensions. The paper's central claim is that SparseMeXt is the first sparse paradigm to surpass dense methods in online HD map construction.","tokens_in":11416,"tokens_out":6929,"duration_ms":66855,"significance":"If the results hold, the paper is a useful demonstration that sparse map construction can be competitive with dense BEV pipelines while remaining efficient, and the component-level analysis identifies concrete design choices worth adopting. The paper's strengths include a structured ablation path from the SparseDrive baseline, same-hardware FPS measurements for several variants, and additional long-range and centerline experiments. The main limitation is that the headline 'surpass dense methods' claim is not currently supported by a controlled comparison at equal backbone capacity, and several component gains are reported from single runs without variance information. These issues are fixable, but they affect the paper's central quantitative narrative.","major_comments":[{"comment":"There is an internal inconsistency in the MapTRv2-Base number: Section IV-B states that SparseMeXt-Base outperforms MapTRv2-Base at 63.0%, while Table VII reports MapTRv2-Base at 68.7%. If Table VII is correct, then SparseMeXt-Base (65.2 mAP) is 3.5 points behind the matched-backbone dense method. This inconsistency directly affects the central claim and must be corrected in both the text and the table.","section":"Section IV-B, Table VII"},{"comment":"The claim that SparseMeXt is the first sparse paradigm to surpass dense methods is not established by the reported comparisons. At the Base/Large scale, the only sparse-over-dense win is SparseMeXt-Large (ResNet-101, 68.9) against MapTRv2-Base (ResNet-50, 68.7), a comparison that is unmatched in backbone capacity; at matched ResNet-50, the dense method is ahead (68.7 vs. 65.2). SparseMeXt-Tiny (ResNet-18, 55.5) does beat MapTRv2-Tiny (ResNet-18, 52.3), so the claim is not vacuous, but a capacity-controlled comparison is needed. Please report a dense baseline trained under the same recipe at matched capacity, or restrict the headline claim to the configuration for which a controlled comparison is available.","section":"Section I, Section IV-B, Table VII"},{"comment":"Several component contributions are supported only by single-run differences of 0.07, 0.6, 0.5, and 0.9 mAP points, with no standard deviation, number of seeds, or significance measure. On nuScenes mAP, differences of this size are commonly within run-to-run noise; as written, the stepwise attribution in Table X and the stage-ratio conclusion in Table III are not firmly supported. Please report at least three seeds with mean and standard deviation for the main ablation path, and re-evaluate whether the 0.07-point stage-ratio gain and the 0.6-point Decouple-DFA gain are reliable.","section":"Tables III, IV, V, VI, X; Section IV-C"},{"comment":"The text repeatedly claims a 10.1% mAP improvement over 'SparseDrive', but the numbers do not line up: Table X's 'SparseDrive-Map' baseline is 57.4, Table VII lists SparseDrive-S at 55.1 and SparseDrive-B at 56.2, and SparseMeXt-Base is 65.2. Depending on the baseline, the absolute gain is 7.8, 9.0, or 10.1 mAP points. Please specify the exact baseline configuration and state clearly whether the reported value is an absolute mAP-point gain or a relative percentage gain.","section":"Sections I and IV-C; Tables VII and X"},{"comment":"PPDN is one of the three central contributions, but Section III-D does not provide enough detail to reproduce the denoising scheme: the rotation, scaling, and curvature noise operations are described verbally, without explicit equations, coordinate conventions, noise magnitude ranges, or sampling probabilities. Please add a precise algorithm or equations for the four noise types and report the hyperparameters used in training.","section":"Section III-D"}],"minor_comments":[{"comment":"The row labeled '1 non-temporal, 4 temporal' displays the same rounded average mAP (57.4) as the baseline but is annotated as '+0.07'; either unrounded values are needed or the annotation should be reconciled with the displayed numbers, and the unit should be mAP points rather than percent.","section":"Table III"},{"comment":"The text says the long-range setting is '60x90m' and '90m perception range', while Table IX labels rows as '90×60m'. Please make the x/y range convention consistent.","section":"Section IV-B / Table IX"},{"comment":"There are numerous typos and inconsistent spellings, including 'TabelX', 'SparseMaXt', 'pdedstrain crossing', 'Nuscenes' vs. 'nuScenes', and 'stage radio' for 'stage ratio'. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"Reference [6] is cited as 'Persformer: A transformer architecture for topological machine learning', but the same name is used in [18] for a 3D lane detection paper; please verify that the correct references are attached to the textual mentions.","section":"Section II-A"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a solid systems paper with genuinely new components and a clean ablation story, but the central marketing claim—first sparse method to surpass dense HD map construction—does not survive a matched-backbone check. The paper's own Table VII shows SparseMeXt-Base (R50) at 65.2 mAP vs MapTRv2-Base (R50) at 68.7. The only sparse-dense win is SparseMeXt-Large (R101) at 68.9 against MapTRv2-Base (R50). That is a capacity difference, not a paradigm win. The text in Section IV-B even misreports MapTRv2-Base as 63.0%, which happens to be MapNeXt-Tiny's number, making the claimed 0.2-point edge look even shakier.\n\nWhat is actually good: the authors give a clear rationale for why sparse detectors underperform on maps—large spatial coverage vs dynamic occupancy—and each design choice is motivated. The SiMo neck alone is worth +4.25 mAP, and nuImages detection pretraining gives +1.5. The decoupled DFA, query-centric sparse-dense segmentation, and physical-prior denoising (PPDN) are new, and the stepwise ablation shows a coherent progression from 57.4 to 65.2. At 20–33 FPS on a 3090, this is practically useful for onboard deployment.\n\nThe soft spots beyond the headline: no error bars or multiple seeds, and several deltas (0.07, 0.5, 0.6, 0.9) are small enough to be noise. The PPDN gain of 0.5 and decouple-DFA gain of 0.6 could vanish with variance. No code is released, which makes replication harder. These are fixable: add seed variance, correct the comparison text, add a dense R101 baseline at matched capacity, and release code.\n\nVerdict: I would send this to peer review—the engineering content is worth engaging with—but the authors need to revise the overclaim and provide more rigorous uncertainty quantification before I would trust the specific numbers. The paper is for the autonomous driving / HD map community, and for readers who care about efficient sparse perception. It deserves a serious referee, not a desk reject, but the referee should push on the comparison table.\n\nRecommendation: major revision with a careful check of the headline claim.","headline":"Useful sparse-HD-map engineering with three new components, but the headline claim of being the first sparse method to beat dense approaches is unsupported by the paper's own matched-backbone numbers.","tokens_in":11984,"tokens_out":3490,"would_cite":false,"duration_ms":31326,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SparseMeXt is claimed to be the first sparse HD-map construction model to surpass dense BEV methods, reaching 68.9% mAP on nuScenes at over 20 FPS.","keywords":["sparse representations","HD map construction","nuScenes","query denoising","auxiliary segmentation","vectorized maps","autonomous driving","single-in-multi-out neck"],"falsifier":"Rerun the ablation sequence in Table X with at least five random seeds under the same training protocol; if the seed-to-seed spread on any component equals or exceeds the reported gain (0.07% stage ratio, 0.6% decouple-DFA, 0.9% segmentation loss, 0.5% PPDN), the stepwise attribution collapses. Alternatively, an independent reproduction of SparseMeXt-Large on the nuScenes validation split could check whether 68.9% mAP actually exceeds MapTRv2-Base's 68.7% beyond chance.","tokens_in":10905,"feed_emoji":"🗺️","tokens_out":5303,"duration_ms":48541,"temperature":0.7,"pith_summary":"Online high-definition map construction has been dominated by dense bird's-eye-view (BEV) representations that are computationally heavy. This paper argues that sparse, query-based representations can close that gap and exceed it, provided the architecture is tailored to map geometry instead of borrowed from 3D object detection. It reports SparseMeXt-Large at 68.9% mAP on nuScenes with over 20 FPS, surpassing the dense MapTRv2-Base at 68.7% and the sparse SparseDrive-B at 56.2%. The paper's claim is that sparse representations are not inherently weaker, and its three design changes are what unlock their potential.","feed_headline":"SparseMeXt hits 68.9% mAP, beating dense map baselines","feed_subtitle":"Sparse HD-map construction surpasses dense MapTRv2 on nuScenes while running over 20 FPS on an RTX 3090.","key_machinery":"The load-bearing mechanism is a sparse encoder-decoder with map-specific modifications at every level. The image encoder uses a nuImages-pretrained ResNet backbone and a single-in-multi-out (SiMo) neck that feeds only the backbone's C5 feature into the decoder, justified by the observation that map elements occupy consistently large image areas and benefit less from multi-scale feature fusion. The decoder splits into independent classification and regression branches at the deformable feature aggregation layer so each task samples its own points. The auxiliary sparse-dense segmentation head converts instance features into a BEV segmentation map during training only, supplying global foreground supervision without dense BEV computation at inference. Finally, physical-prior query denoising (PPDN) injects rotation, location, scale, and curvature perturbations that respect the geometry of polylines and polygons, stabilizing DETR-style query training for point-sequence outputs.","core_discovery":"The central discovery is that sparse HD map construction can match and surpass dense BEV methods when each design stage is revisited for map-specific properties. SparseMeXt achieves this through coordinated changes: a backbone pre-trained on driving-domain detection rather than ImageNet classification, replacing the multi-in-multi-out FPN neck with a single-in-multi-out neck because map elements are large objects, adjusting the balance between non-temporal and temporal decoder stages, and decoupling classification from regression in the deformable feature aggregation. On top of the architecture, the paper adds a query-centric sparse-dense auxiliary segmentation loss for global foreground supervision, which costs nothing at inference, and a physical-prior query denoising scheme with rotation, location, scale, and curvature noise tailored to line-shaped map elements. With these components, SparseMeXt-Base reaches 65.2% mAP and SparseMeXt-Large reaches 68.9% mAP, the latter exceeding the dense MapTRv2-Base by 0.2 points while running at over 20 FPS on an RTX 3090.","pith_inferences":["Beyond the paper: the SiMo neck result suggests that large-object perception tasks may not need multi-scale fusion at all; lane detection and drivable-area segmentation are natural test beds.","Beyond the paper: the physical-prior noise modes (rotation, shift, scale, curvature) are geometry-agnostic and could be applied to trajectory prediction or any polyline regression, where box-style noise is a poor fit.","Beyond the paper: the +1.5% mAP attributed to nuImages pre-training is the most portable single finding; a multi-seed study across backbones would tell whether task-relevant pre-training, not architecture, drives the gain."],"forward_implications":["If the 68.9% mAP result holds, sparse map construction is no longer a compromise: it beats the dense state-of-the-art MapTRv2-Base while avoiding dense BEV features, at over 20 FPS.","The design principle generalizes: because map elements are large and static, a single-scale neck, a 1+4 stage ratio, and decoupled sampling can be chosen by that geometry rather than by 3D detection defaults.","The auxiliary segmentation head and PPDN add accuracy at training time only, so the deployed model keeps the sparse pipeline's speed and memory advantage.","Centerline prediction extends the same architecture to 58.8% mAP on nuScenes, connecting map construction to downstream motion prediction and planning.","Long-range perception to 60 m by 90 m is feasible with a sparse model, where SparseMeXt-Base outscores MapTR by 7.4% mAP."],"supporting_citations":[{"why":"Supplies the fully sparse SparseDrive baseline and training protocol that SparseMeXt starts from and improves by 10.1% mAP.","marker":"[17]"},{"why":"Dense MapTRv2 baseline that SparseMeXt-Large surpasses (68.9 vs 68.7) and source of the auxiliary foreground segmentation idea that the paper adapts to a sparse-dense form.","marker":"[15]"},{"why":"MapTR defines the DETR-style vectorized map paradigm used for comparison and provides the long-range MapTR baseline in Table IX.","marker":"[14]"},{"why":"DN-DETR supplies the query denoising concept that physical-prior query denoising extends to point-sequence map elements.","marker":"[20]"},{"why":"YOLOF provides the single-in-multi-out neck and the divide-and-conquer analysis behind the SiMo encoder neck design.","marker":"[35]"},{"why":"Provides the analysis of FPN back-propagation paths and MiMo structure that motivates replacing the multi-scale neck.","marker":"[34]"},{"why":"MapNeXt is a dense strong baseline (63.0% mAP) that SparseMeXt surpasses and a source of scaling practices.","marker":"[36]"}],"fun_headline_variants":["SparseMeXt hits 68.9% mAP, outpaces dense HD mapping","Sparse representation beats dense in HD maps at 68.9%","Sparse map construction beats dense at 68.9% mAP","Sparse HD map hits 68.9% mAP, surpassing dense methods","SparseMeXt: sparse beats dense with 68.9% mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ablation deltas of 0.07 to 0.9 mAP points reported in Section IV-C are real improvements rather than single-run noise; the paper does not report variance or multiple seeds for those steps.","fun_headline_variants_meta":{"raw":{"variants":["SparseMeXt hits 68.9% mAP, outpaces dense HD mapping","Sparse representation beats dense in HD maps at 68.9%","Sparse map construction beats dense at 68.9% mAP","Sparse HD map hits 68.9% mAP, surpassing dense methods","SparseMeXt: sparse beats dense with 68.9% mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000654,"raw_usage":{"total_tokens":3057,"prompt_tokens":1066,"completion_tokens":1991,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":682,"completion_tokens_details":{"reasoning_tokens":1887}},"tokens_in":682,"tokens_out":1991,"duration_ms":12897,"temperature":1.0,"reasoning_tokens":1887,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:22:21.663965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the ablation sequence in Table X with at least five random seeds under the same training protocol; if the seed-to-seed spread on any component equals or exceeds the reported gain (0.07% stage ratio, 0.6% decouple-DFA, 0.9% segmentation loss, 0.5% PPDN), the stepwise attribution collapses. Alternatively, an independent reproduction of SparseMeXt-Large on the nuScenes validation split could check whether 68.9% mAP actually exceeds MapTRv2-Base's 68.7% beyond chance.","supporting_citations":[{"cited_title":"Dn- detr: Accelerate detr training by introducing query denoising,","cited_arxiv_id":null,"evidence_quote":"DN-DETR supplies the query denoising concept that physical-prior query denoising extends to point-sequence map elements."},{"cited_title":"You Only Look One-level Feature","cited_arxiv_id":"2103.09460","evidence_quote":"YOLOF provides the single-in-multi-out neck and the divide-and-conquer analysis behind the SiMo encoder neck design."},{"cited_title":"You Should Look at All Objects","cited_arxiv_id":"2207.07889","evidence_quote":"Provides the analysis of FPN back-propagation paths and MiMo structure that motivates replacing the multi-scale neck."}],"review_version":1}