{"id":"2b56f869-408d-497b-bccf-57bb72b2d4fa","arxiv_id":"2507.21423","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MapDiffusion uses a diffusion decoder conditioned on a latent BEV grid to sample multiple vectorized HD maps, reporting a 5.3% relative mAP gain over StreamMapNet on nuScenes and uncertainty that increases in occluded areas.","lead":"MapDiffusion applies diffusion models to online HD map construction in autonomous driving, generating multiple plausible vectorized maps instead of one deterministic prediction. The paper also derives spatial uncertainty from the spread of sampled maps, which could help self-driving systems recognize ambiguous scenes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The uncertainty claim is not calibrated against map error; the visibility-correlation test in §IV-E.2 is indirect, so the 'useful spatial uncertainty signal' half of the central claim lacks direct support.","rationale":"The Reader's verdict (CONDITIONAL) and rationale already note the non-reproduced StreamMapNet baseline and the absence of error bars/code. That is a real concern, but the most load-bearing part of the central claim is the uncertainty signal: it is the paper's stated novelty beyond the accuracy gain. The evidence in §IV-E.2 is a single aggregate visibility comparison with a t-test. It demonstrates a correlation with a visibility proxy, not with map error, and no control baseline is tested. The proposed calibration experiment would settle whether U is genuinely useful for downstream decisions. If it fails, the paper's main contribution is reduced to an incremental mAP gain on one split; if it passes, the conditional accept is justified. I therefore keep the Reader's CONDITIONAL verdict unchanged rather than moving to ACCEPT or REJECT.","tokens_in":12770,"tokens_out":8550,"duration_ms":104139,"concrete_test":"Compute the n=10 aggregated map D_c from §IV-E.1 and binarize it with the ROC-optimal threshold; build a binary error map E(x,y) = 1 if D_c(x,y) disagrees with the rasterized GT within a small tolerance. Then measure the AUROC of U(x,y) (Eq. 4) as a predictor of E(x,y), or the Spearman correlation between U and a soft error (e.g., |D_c - GT|). If AUROC is near 0.5 on held-out scenes, the uncertainty claim fails. As a control, run the same analysis with StreamMapNet plus independent Gaussian noise on output polylines with variance matched to MapDiffusion's sample variance; if the control achieves similar AUROC, variance is not specifically informative.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two load-bearing parts: the 5.3% accuracy gain and the usefulness of sample variance as uncertainty. The accuracy part rests on an asterisked, non-reproduced StreamMapNet baseline (Table I), which the Reader already flags. The more directly load-bearing weakness is the uncertainty part, because it is the paper's primary novelty. Section IV-E.2 compares mean uncertainty U in visible vs. invisible areas (Fig. 5: 0.0063 vs. 0.0082, 31% higher, p<0.001). This establishes only that U correlates with a visibility proxy, not that U predicts where the map is wrong. A stochastic baseline whose noise is larger in occluded regions would pass this test without providing useful uncertainty. Moreover, U is computed in raster space from Gaussian-smoothed, clipped class probabilities (Eqs. 1-4), so sample variance may reflect rasterization artifacts rather than geometric ambiguity. No calibration experiment (e.g., does high U coincide with high per-pixel error against GT?) and no comparison to Monte Carlo dropout or random query perturbation is provided. If U does not track map error, the 'uncertainty-aware decision-making' contribution in the abstract and conclusion is unsupported, even if the mAP gain survives.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"MapDiffusion proposes to replace the deterministic DETR-style query decoder of StreamMapNet with a diffusion-based decoder that iteratively denoises randomly initialized polylines conditioned on a BEV latent grid. The model is trained with a noise scheduler on vectorized map elements and sampled with DDIM, producing multiple map hypotheses. The paper reports 35.6% mAP on the nuScenes split (vs. an asterisked 33.8% for StreamMapNet), shows that aggregating 10 rasterized samples improves ROC AUC from 0.89 to 0.92, and computes a spatial uncertainty map U from cross-sample variance, which is 31% higher in invisible than in visible areas (p<0.001). Ablations cover diffusion steps, DDIM eta, query threshold, BEV encoder pretraining, and query padding strategies.","tokens_in":13053,"tokens_out":3739,"duration_ms":48405,"significance":"If the accuracy and uncertainty claims hold, MapDiffusion is a useful demonstration that a full generative diffusion process can operate directly on vectorized map elements for online HD map construction, rather than on raster maps or as a refinement of existing proposals. The paper is explicit in comparing against PolyDiffuse and positions the uncertainty estimate as the main novelty. The ablations are thorough regarding diffusion parameters and padding strategies, and the sample-aggregation experiment provides a concrete check that multiple samples carry information. However, the current evidence for the headline accuracy gain relies on a non-reproduced baseline, and the uncertainty validation only shows a correlation with a visibility proxy rather than with map error, leaving the decision-relevant claim under-supported. The paper's strengths are its clear formulation of the diffusion setup over polylines and the recognition that sample variance is a natural uncertainty signal.","major_comments":[{"comment":"The central accuracy claim (35.6% vs. 33.8% mAP, a 5.3% relative improvement) is computed against an asterisked StreamMapNet baseline taken from the original paper [8], not reproduced with the same training pipeline, evaluation code, or hardware. Because MapDiffusion changes the decoder, query initialization, and training procedure, the comparability of the two numbers is not established. Please reproduce StreamMapNet under identical conditions and, for both models, report mean and standard deviation over multiple seeds; the mAP differences in Table II (35.2 to 35.6) are small enough that single-run variability could affect the conclusion.","section":"Section IV-D, Table I"},{"comment":"The uncertainty evaluation demonstrates only that U is higher in invisible/occluded areas (mean 0.0082 vs. 0.0063, 31% higher, p<0.001). It does not test whether U correlates with map prediction error, which is the decision-relevant quantity claimed in the abstract and conclusion. A stochastic model whose output noise is larger in occluded regions would pass this test without providing useful uncertainty for planning. Please add a direct calibration analysis (for example, bin U and measure per-pixel or per-instance error against GT maps), compare against Monte Carlo dropout or random query perturbation, and preferably evaluate a downstream use as in Gu et al. [30]. Without such evidence, the phrase 'uncertainty-aware decision-making' in the conclusion is not supported.","section":"Section IV-E.2, Eq. (4)"},{"comment":"The aggregation benefit is demonstrated only in raster space, using the ROC/AUC of rasterized probability maps (Eqs. 1-3). The output representation used by planning systems is the vectorized map, and the paper does not show that aggregating samples improves vectorized instance-level metrics such as mAP or the quality of the final polyline set. Since Section III-C.1 states that raster-space aggregation is 'sufficient to demonstrate our point', please make explicit that the claimed accuracy improvement from aggregation is restricted to rasterized evaluation, or provide vectorized aggregation results.","section":"Section IV-E.1, Section III-C.1"},{"comment":"The model reaches 35.2% mAP with a single diffusion step and 35.6% with five steps, so the accuracy gain from the iterative diffusion process itself is small. Since the StreamMapNet baseline also uses iterative refinement in its decoder, the specific contribution of the diffusion mechanism over a deterministic refinement scheme is not isolated. Please add an ablation with the same decoder but deterministic refinement or with learned queries, and report sample-diversity statistics (e.g., pairwise distance between sampled maps) to substantiate that the model is generating a meaningful distribution rather than noise around a single mode.","section":"Section IV-G.1, Table II"}],"minor_comments":[{"comment":"The sentence 'Table I shows the qualitative results' should read 'quantitative results'.","section":"Section IV-D"},{"comment":"Please clarify the unit of analysis for the t-test: the text says 'mean uncertainty per traffic scene' but the figure shows a distribution of variances; if the test is per-pixel, spatial autocorrelation would make the p-value invalid.","section":"Section IV-E.2"},{"comment":"The 'Smooth' padding strategy is described only as 'smooth random polylines (both straight and curved) or polygons'; the generation procedure should be specified or referenced.","section":"Section IV-G.3"},{"comment":"Inference FPS is reported on an NVIDIA A10 while training is on V100s; please state the GPU used for each FPS number and whether timing includes the BEV encoder.","section":"Section IV-B"},{"comment":"The spatial resolution H x W of the uncertainty map U is not specified; please state whether it matches the BEV grid (100x50) or another resolution.","section":"Section III-C.2, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a computer vision venue, but the uncertainty claim is the paper's distinctive novelty and currently rests on an indirect visibility correlation. The lack of a reproduced StreamMapNet baseline and the absence of seed statistics make the headline 5.3% gain fragile. I would encourage the editors to require the calibration experiment and the baseline rerun before publication; without those, the paper's central promises are only partially supported. The raster-only aggregation result is also more limited than the abstract implies."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MapDiffusion is a genuine new entry in online vectorized HD mapping: it runs a DDIM denoiser over polyline queries conditioned on a BEV latent, generating full maps from noise rather than refining coarse proposals like PolyDiffuse. The 35.6% mAP on the nuScenes split, if the asterisked StreamMapNet baseline of 33.8% holds up, is a decent 5.3% relative gain. The ablations on k, eta, and tau are careful, and the padding strategy study is sensible. The authors also made a real effort to check their uncertainty signal externally: they built visibility masks from Occ3D and showed sample variance is 31% higher in occluded areas, with a t-test p<0.001.\n\nThe soft spots are proportional. The baseline is taken from the StreamMapNet paper, not rerun; the authors do rerun SQD-MapNet, so they know how to run baselines, but not the primary one. No error bars or multi-seed numbers appear anywhere, so the headline 35.6 vs 33.8 could move with seeds. The diffusion component is not isolated: a DETR-style decoder with learned queries and the same iterative refinement but no noise would clarify what diffusion itself adds. The aggregation benefit is shown only in raster space, not in the vectorized mAP metric, so the claim that distribution modeling helps prediction is demonstrated on a proxy task.\n\nThe uncertainty claim is the one I want a referee to press on. The visibility test shows U increases where the camera is occluded. That is necessary but not sufficient for a useful uncertainty map: a stochastic baseline whose noise is larger in occluded regions regardless of error would pass. The paper frames U as a qualitative indicator, so it does not overclaim calibration, but the conclusion calls these \"useful uncertainty maps\" and the abstract promises \"uncertainty-aware decision-making.\" Without a calibration check against map error, or at least a comparison to Monte Carlo dropout, that half of the central claim is suggestive, not established.\n\nOverall, the paper is clearly written, the experiments are reproducible in principle, and the authors are transparent about what they did and did not do. The weaknesses are standard for this area: missing code, missing seeds, a borrowed baseline. None of them is fatal. A referee with a reasonable request list can turn this into a solid publication.\n\nThis paper is for researchers working on online HD map construction, uncertainty in perception for autonomous driving, and diffusion for structured prediction. It deserves a serious referee; I would send it to review with requests for a reproduced baseline, error bars, a calibration experiment, and code.","headline":"A solid, well-engineered application of diffusion to vectorized HD map construction; the mAP gain is plausible but rests on a non-reproduced baseline, and the uncertainty story is suggestive rather than proven.","tokens_in":13563,"tokens_out":2421,"would_cite":true,"duration_ms":27701,"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":"MapDiffusion's claim: full generative diffusion over vectorized polylines beats the StreamMapNet baseline and turns sample spread into an uncertainty signal.","keywords":["online HD map construction","diffusion models","vectorized map","uncertainty estimation","bird's-eye view","autonomous driving","nuScenes","DDIM sampling"],"falsifier":"Rerun StreamMapNet with the identical training recipe and split used for MapDiffusion; if its reproduced mAP reaches or exceeds 35.6%, the claimed 5.3% relative improvement and the state-of-the-art conclusion are unsupported. A second check is to apply the variance-based uncertainty map to scenes with no occlusion; if variance remains high on fully visible geometry, then sample spread is not tracking scene ambiguity.","tokens_in":12606,"feed_emoji":"🗺️","tokens_out":7021,"duration_ms":75409,"temperature":0.7,"pith_summary":"The paper tries to establish that online HD map construction can be formulated as a generative diffusion problem rather than a deterministic prediction problem. MapDiffusion denoises randomly initialized polylines conditioned on a BEV latent grid, so one forward pass yields a sample from the map distribution instead of a single committed map. The experiments report 35.6% mAP versus 33.8% for StreamMapNet, a 5.3% relative gain, at 8.0 FPS with five diffusion steps, and improved ROC AUC from 0.89 to 0.92 when ten samples are aggregated. The same samples give a spatial uncertainty estimate whose mean is 31% higher in occluded regions, suggesting that ambiguity is quantified where it matters. This matters because a self-driving planner can use both the map distribution and its uncertainty rather than trusting one deterministic interpretation.","feed_headline":"Diffusion over polylines lifts online HD map mAP to 35.6 percent","feed_subtitle":"Sampling many maps beats one prediction and flags occluded areas with 31 percent higher variance.","key_machinery":"The central mechanism is the graph diffusion decoder: a DETR-style transformer decoder turned into a denoiser. It starts from randomly initialized polylines, takes a diffusion time-step embedding, and queries the latent BEV grid through deformable cross-attention; a cosine noise scheduler defines the forward process, and training minimizes a line reconstruction loss plus a class loss. This mechanism is what converts map construction from a single deterministic prediction into sampling from a map distribution, and the same sampled polylines feed both the aggregated prediction and the variance-based uncertainty map.","core_discovery":"MapDiffusion's central claim is that full generative diffusion over vectorized polylines, conditioned on a BEV latent grid, is a viable replacement for deterministic vector decoding in online HD map construction, and that the resulting sample distribution carries a usable uncertainty signal. On the nuScenes split without geospatial overlap, the single-sample prediction reaches 35.6% mAP against 33.8% for StreamMapNet, a 5.3% relative gain, at 8.0 FPS with five DDIM steps. Sampling ten maps and aggregating them in raster space improves ROC AUC from 0.89 to 0.92. The per-location variance across samples is the uncertainty estimate, and it is 31% higher in invisible than in visible areas, a difference the paper reports as significant at p < 0.001.","pith_inferences":["The paper does not evaluate downstream planning, but feeding the per-location variance into a trajectory predictor or planner is a direct testable extension; higher variance in occluded areas should reduce confidence in any single map-based plan.","Because the BEV encoder is computed once and only the decoder is iterated, the same diffusion decoder could be attached to newer non-diffusion map heads; if so, those models would inherit sample diversity and uncertainty without redesigning their encoder.","The 31% gap between visible and invisible uncertainty likely understates the effect of occlusion, since the StreamMapNet-style BEV temporal aggregation leaks information from previous frames into currently occluded areas; a single-frame variant should show a larger gap.","A useful stress test is to check whether sample variance in occluded regions reflects genuine topological ambiguity, such as one lane versus two, rather than just noise; scoring samples by distinct topologies would separate those cases."],"forward_implications":["Aggregating 10 rasterized samples raises ROC AUC from 0.89 to 0.92 relative to a single sample, so sampling is an accuracy lever, not just a visualization tool.","The variance across samples is significantly higher in invisible areas, with mean 0.0082 versus 0.0063, giving a spatial uncertainty signal that tracks occlusion.","MapDiffusion reaches 35.6% mAP without learned queries or decoder-side temporal aggregation, so the generative decoder can replace the query mechanism rather than merely refine it.","Because the BEV encoder runs once, inference stays real-time: 8.0 FPS with five diffusion steps and 12.8 FPS with one step on the reported hardware.","Class prediction is excluded from the diffusion loop, so denoising happens purely in vector space, which is what lets the model generate vectorized maps directly instead of raster masks."],"supporting_citations":[{"why":"Supplies the reference architecture, the nuScenes split without geospatial overlap, and the 33.8% mAP baseline against which the headline gain is measured.","marker":"[8]"},{"why":"Provides the Denoising Diffusion Probabilistic Models formulation that MapDiffusion adapts to vector polyline space.","marker":"[15]"},{"why":"Provides the DDIM sampling schedule that lets MapDiffusion generate high-quality samples with only five diffusion steps.","marker":"[36]"},{"why":"Supplies the DETR-style transformer decoder structure that MapDiffusion repurposes as its denoising decoder.","marker":"[34]"},{"why":"Supplies the deformable cross-attention mechanism used to condition the denoising decoder on the latent BEV grid.","marker":"[35]"},{"why":"Supplies the SQD-MapNet query-denoising baseline reproduced on the same split at 33.1% mAP, which MapDiffusion must beat.","marker":"[14]"},{"why":"PolyDiffuse is the only prior diffusion-based vectorized map method, and MapDiffusion distinguishes itself as full generation rather than refinement of coarse proposals.","marker":"[11]"},{"why":"Supplies the nuScenes dataset used for all training and evaluation.","marker":"[37]"},{"why":"Supplies the Occ3D occupancy maps from which the paper derives visibility masks for the occlusion-uncertainty evaluation.","marker":"[42]"}],"fun_headline_variants":["MapDiffusion: generative sampling boosts HD map accuracy and uncertainty","Diffusion-based HD map construction: sample to improve and estimate uncertainty","MapDiffusion: 5% better HD maps via diffusion sampling, uncertainty estimates","HD map generation via diffusion: multiple samples beat deterministic output","MapDiffusion: sampling maps for HD construction improves mAP and uncertainty"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 33.8% mAP StreamMapNet baseline is taken from the original paper and was not reproduced under identical training conditions; the claimed 5.3% improvement depends on that number being directly comparable to the MapDiffusion run, and if the baseline were rerun and scored higher, the headline gain would shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["MapDiffusion: generative sampling boosts HD map accuracy and uncertainty","Diffusion-based HD map construction: sample to improve and estimate uncertainty","MapDiffusion: 5% better HD maps via diffusion sampling, uncertainty estimates","HD map generation via diffusion: multiple samples beat deterministic output","MapDiffusion: sampling maps for HD construction improves mAP and uncertainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1485,"prompt_tokens":979,"completion_tokens":506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":414}},"tokens_in":595,"tokens_out":506,"duration_ms":6625,"temperature":1.0,"reasoning_tokens":414,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:46:21.702115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun StreamMapNet with the identical training recipe and split used for MapDiffusion; if its reproduced mAP reaches or exceeds 35.6%, the claimed 5.3% relative improvement and the state-of-the-art conclusion are unsupported. A second check is to apply the variance-based uncertainty map to scenes with no occlusion; if variance remains high on fully visible geometry, then sample spread is not tracking scene ambiguity.","supporting_citations":[{"cited_title":"Streammapnet: Streaming mapping network for vectorized online hd map construc- tion,","cited_arxiv_id":null,"evidence_quote":"Supplies the reference architecture, the nuScenes split without geospatial overlap, and the 33.8% mAP baseline against which the headline gain is measured."},{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Provides the Denoising Diffusion Probabilistic Models formulation that MapDiffusion adapts to vector polyline space."},{"cited_title":"Denoising diffusion implicit models,","cited_arxiv_id":null,"evidence_quote":"Provides the DDIM sampling schedule that lets MapDiffusion generate high-quality samples with only five diffusion steps."},{"cited_title":"End-to-end object detection with transformers,","cited_arxiv_id":null,"evidence_quote":"Supplies the DETR-style transformer decoder structure that MapDiffusion repurposes as its denoising decoder."},{"cited_title":"Deformable DETR: Deformable transformers for end-to-end object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the deformable cross-attention mechanism used to condition the denoising decoder on the latent BEV grid."},{"cited_title":"Stream query denoising for vectorized hd- map construction,","cited_arxiv_id":null,"evidence_quote":"Supplies the SQD-MapNet query-denoising baseline reproduced on the same split at 33.1% mAP, which MapDiffusion must beat."},{"cited_title":"Polydiffuse: Polygonal shape reconstruction via guided set diffusion models,","cited_arxiv_id":null,"evidence_quote":"PolyDiffuse is the only prior diffusion-based vectorized map method, and MapDiffusion distinguishes itself as full generation rather than refinement of coarse proposals."},{"cited_title":"nuScenes: A multimodal dataset for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Supplies the nuScenes dataset used for all training and evaluation."},{"cited_title":"Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Supplies the Occ3D occupancy maps from which the paper derives visibility masks for the occlusion-uncertainty evaluation."}],"review_version":1}