{"id":"7ee4998b-c053-4ad7-8a2f-edacd414ef20","arxiv_id":"2411.17161","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using crowdsourced trajectory heatmaps and vectorized tokens as fusion priors improves lane segment mAP from 32.30 to 42.30 and topology score from 25.40 to 34.65 on OpenLane-V2.","lead":"This paper adds historical crowdsourced vehicle trajectories as a map prior for an online lane perception model, alongside standard-definition maps, and reports large gains on the OpenLane-V2 benchmark. The idea is practical: trajectory data is cheaper and closer to true lane geometry than coarse SD maps.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I's +10.0/+9.25 SOTA claim is not apples-to-apples: TrajTopo may be scored on a trajectory-density-filtered validation subset using priors from geographically overlapping Argoverse2 splits, and the clean-split check omits all prior SOTA baselines.","rationale":"The reader's weakest assumption correctly identifies geographic overlap and the incomplete clean-split comparison as the central risk. I agree, and I sharpen the concern by adding a second protocol mismatch: the trajectory-density frame filter described in Section V.A.1 may make the validation set used for TrajTopo easier than the full set used for prior SOTA numbers. In good faith, the paper does include a new-split check, and the trajectory prior still helps there relative to LaneSegNet, so the direction is credible. However, the headline 'surpasses state-of-the-art' claim is not established until the evaluation subsets and baselines are matched. This does not change the reader's conditional verdict: the core contribution is plausible but requires released artifacts and controlled comparison before the specific numerical gains can be taken at face value.","tokens_in":11426,"tokens_out":6452,"duration_ms":61827,"concrete_test":"Use the released code and OpenLane-V2 to run a matched-protocol comparison: evaluate TrajTopo, LaneSegNet, P-MapNet, SMERF, and MapEx on the same full validation set with no frame filtering; then evaluate at least SMERF and MapEx on the exact filtered subset used for TrajTopo; finally add SMERF/MapEx/P-MapNet baselines to the StreamMapNet new-split rows. If the +10.0/+9.25 margins shrink to about 2 points or less on the full set, or if the new-split SOTA baselines match TrajTopo, the 'outperforms SOTA' claim is an artifact of evaluation protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on Table I's comparison, but that comparison is not controlled. Section V.A.1 filters frames to those with more than five times as many trajectories as centerlines; if applied at evaluation, TrajTopo is reported on the roughly 86% of validation frames with abundant trajectories (Section III), whereas published P-MapNet, SMERF, and MapEx numbers are conventionally computed on the full OpenLane-V2 validation set. A model evaluated on an easier, trajectory-rich subset would show inflated gains even without any prior. Second, the prior itself is drawn from Argoverse2 motion-forecasting logs; the paper itself cites a 54% geographic overlap between Argoverse2 train and validation splits. On the standard split, trajectory priors can therefore encode lane geometry in the very neighborhoods being evaluated, which is closer to memorized map look-up than to online perception. The new-split evaluation is the right control, but Table I's lower rows only list LaneSegNet* and TrajTopo* — no SMERF*, MapEx*, or P-MapNet* — so it does not establish 'surpassing SOTA' under a clean protocol. The claimed +10.0/+9.25 margin is not independently checkable until the exact validation frame IDs, the filtered subset, and the trajectory-source mapping are released.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TrajTopo, a method that augments lane segment perception and topology reasoning on the OpenLane-V2 benchmark with crowdsourced trajectory priors extracted from the Argoverse2 motion forecasting dataset. Trajectories are encoded as a rasterized heatmap or as vectorized instance tokens via K-means or farthest-point sampling, and are fused into a BEVFormer-style encoder and a DETR-style decoder. A confidence-based fusion module with spatial alignment and BEV segmentation supervision is introduced to mitigate prior-to-perception misalignment. The authors report that their best configuration improves lane segment mAP by +10.0 and topology metric by +9.25 over prior SOTA methods, and they also include a new-split evaluation intended to control for geographic overlap between Argoverse2 train and validation data.","tokens_in":11654,"tokens_out":6094,"duration_ms":54700,"significance":"The central idea of using crowdsourced trajectories as a map prior is interesting and potentially practical: trajectory data are broadly available, frequently updated, and closely reflect lane-level geometry, as the paper's own IoU analysis suggests. The within-framework ablation in Table II, which compares SDMap and trajectory priors on top of the same LaneSegNet baseline, is a genuinely controlled piece of evidence and supports the conclusion that trajectory priors help. The paper also releases code and includes a new-split robustness check, which is a good-faith attempt to address known data-leakage concerns in this benchmark family. However, the headline SOTA claim is not supported by the evidence as presented: the comparison in Table I stacks multiple engineering changes with the trajectory module, the evaluation subset is defined using a trajectory-density filter that is not clearly applied to baselines, and the new-split evaluation omits all prior SOTA methods. If these evaluation issues are fixed, the contribution would be a solid empirical study of a useful prior modality.","major_comments":[{"comment":"The headline comparison (+10.0/+9.25 over prior methods) is not controlled. TrajTopo includes the LaneDAG topology head, an SDMap fusion module, trajectory fusion, and the alignment module, whereas P-MapNet, SMERF, and MapEx each use their own backbones and fusion designs. The text says the authors 'reproduce the state-of-the-art SDMap prior fusion algorithm based on the LaneSegNet framework,' but the details of this reproduction (which method, which hyperparameters, how the LaneDAG head is combined) are not given, so the reader cannot verify that the comparison isolates the trajectory prior rather than the accumulated engineering changes. Please report the SOTA baselines under the same framework and training recipe, or clearly present Table I as a system-level comparison and Table II as the controlled ablation.","section":"Section V.B, Table I"},{"comment":"The filtering statement 'we retain only the frames where the number of trajectories exceeded five times the number of centerlines' is ambiguous. If this filter is applied to the evaluation set, TrajTopo is scored on the approximately 86% of validation frames richest in trajectory coverage while the published P-MapNet, SMERF, and MapEx numbers are computed on the full validation set, which would inflate the reported gains independently of any perceptual improvement. Please state explicitly whether the filter is applied during training, evaluation, or both; if evaluation is filtered, also report results on the full validation set and release the exact frame IDs used so the comparison is reproducible.","section":"Sections III and V.A.1"},{"comment":"The new-split evaluation, which is the only control for the 54% geographic overlap between Argoverse2 train and validation splits, compares only LaneSegNet* and TrajTopo*. It does not include P-MapNet*, SMERF*, MapEx*, or any other SDMap-prior method, so it cannot support the abstract's claim that the method 'significantly outperforms the current state-of-the-art methods' under a leakage-free protocol. Please run the full set of baselines on the same new split, or explicitly limit the claim to the comparison against LaneSegNet.","section":"Section V.B, new-split rows"},{"comment":"Even with the new-split control, the standard-split numbers are the primary evidence for the +10.0/+9.25 claim, and those numbers are obtained with trajectory priors harvested from the same geographic areas as the OpenLane-V2 validation set. The paper acknowledges the 54% overlap but does not quantify how much of the gain comes from memorizing lane geometry in previously seen cities. A stronger test would be to evaluate on the new split with all SOTA baselines, and additionally to ablate the prior by removing trajectory data from areas overlapping the training split; without this, the magnitude of the claimed improvement over SOTA remains uncertain.","section":"Section III and V.B"}],"minor_comments":[{"comment":"The text contains several typos, including 'Code is released is at', 'forcasting', 'Rasterzied', 'ues', and 'askin to'; please revise.","section":"Abstract and throughout"},{"comment":"The definitions of AEtype and AEdist are terse; please specify the matching procedure used to pair predictions with ground truth, since these are not standard OpenLane-V2 metrics and the reader cannot otherwise compare with prior work.","section":"Section V.A.2"},{"comment":"The K-means implementation does not specify the number of clusters K, the trajectory length n used to form A*, or the coverage threshold for retaining frames; these are needed for reproducibility.","section":"Section IV.B.2"},{"comment":"The column headers 'Heatmap', 'Far Sampling', and 'KMeans Cluster' and the row labels are unclear; 'Far Sampling' should be 'Farthest Point Sampling' and the table should state which encoding is used in each fusion strategy.","section":"Table III"},{"comment":"The centerline-aligned IoU computation for SDMap and trajectory data is described in one sentence; please report the rendering resolution, line width, and matching threshold so the IoU values 0.16 and 0.39 are reproducible.","section":"Section III"},{"comment":"No error bars or repeated-seed statistics are reported for any of the main results; given that the margins between some ablations are small (e.g., Table V), at least the key comparisons should include variance over seeds.","section":"Section V.B"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is worth pursuing, and the within-framework ablation plus the new-split check show good faith. The main barrier to acceptance is the evaluation protocol: the SOTA comparison needs to be made apples-to-apples, the trajectory-density filter must be clarified, and the new-split numbers should be provided for all baselines. I would be willing to review a revision that addresses these points; the code release is a positive factor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid engineering paper with a genuinely new idea — using crowdsourced trajectories as a lane-structure prior for online HD mapping — and the authors do a good-faith robustness check. The central direction holds up, but the headline SOTA numbers are not yet supported by a controlled comparison.\n\nWhat's new: prior work uses SD maps, satellite images, or temporal memory; Autograph uses trajectories but offline with aerial imagery. Here the prior is online, aligned via vehicle pose, and encoded as a heatmap or vectorized tokens. The core ablation (Table II) shows that trajectory alone beats SDMap within the same framework (+6.8 vs +5.0 APls over baseline), and all fusion variants help (Table III). That's a real result.\n\nThey also deserve credit for acknowledging the 54% geographic overlap between Argoverse2 train/val and for running a new-split evaluation. That is the right control, and TrajTopo* still improves over its own baseline (16.67 -> 22.50 APls). So the idea is not a mirage.\n\nThe soft spots are real but proportionate. First, Table I's SOTA comparison stacks LaneDAG head, SDMap, and the alignment module on top of the trajectory module, so the +10.0/+9.25 margin is not a clean attribution. Second, Section V.A.1 filters frames to those with >5x trajectories as centerlines; if that filter is applied when reporting Table I, TrajTopo is evaluated on ~86% of the validation frames while the published SMERF/MapEx/P-MapNet numbers are on the full set. That alone could inflate the gain. Third, the new-split table omits those SOTA baselines, so 'surpassing SOTA' is never tested on clean data. Finally, no error bars or multiple runs are reported; differences of a few points at this scale may not be significant.\n\nThe paper is worth a serious referee. The authors should be asked to release the trajectory dataset, the exact validation frame IDs, and to run the SOTA baselines on the new split. If those numbers hold, this is a practical advance. For now, treat the headline as conditional.","headline":"Real idea — trajectory priors for lane perception — but the +10.0 SOTA claim is not yet clean; worth a serious referee and a request for data.","tokens_in":12253,"tokens_out":2965,"would_cite":false,"duration_ms":26536,"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":"Using crowdsourced vehicle traces as map priors lifts lane-segment mAP by 10.0 and topology reasoning by 9.25 on OpenLane-V2.","keywords":["lane segment perception","topology reasoning","trajectory prior","crowdsourcing data","online HD map construction","BEV fusion","OpenLane-V2","Argoverse2"],"falsifier":"Re-run the full comparison on a split where the trajectory-collection and evaluation areas are strictly disjoint (for example, trajectories from one set of cities and perception frames from another); if the trajectory prior then gives no consistent improvement over the SD-map baseline, the reported gains are explained by geographic overlap rather than by the lane-structure content of the priors.","tokens_in":11170,"feed_emoji":"🛣️","tokens_out":6302,"duration_ms":51901,"temperature":0.7,"pith_summary":"This paper argues that crowdsourced vehicle trajectories, collected from the Argoverse2 motion-forecasting dataset, can serve as a lane-structure prior that is more accurate than standard-definition maps, and that injecting them into an online lane-perception model materially improves both lane-segment detection and topology reasoning. The authors encode trajectories in two forms, a rasterized density-and-direction heatmap and vectorized instance tokens from clustering or farthest-point sampling, then fuse them into a BEVFormer-style encoder and a DETR-style decoder. To handle the inevitable spatial and semantic misalignment between prior traces and online camera perception, they add an alignment module that predicts per-cell offsets, warps the prior, and fuses it with confidence weights under BEV segmentation supervision. On OpenLane-V2, the best configuration raises lane-segment average precision by 10.0 and topology score by 9.25 over a strong SD-map prior baseline, and the model also retains an advantage when re-evaluated under a new split designed to limit geographic overlap. A sympathetic reader would take the paper as establishing trajectory data as a practical, easily updated prior for online HD map construction.","feed_headline":"Trajectory priors lift lane mAP by 10.0 and topology by 9.25","feed_subtitle":"Driver traces from Argoverse2, aligned and fused into a lane-perception model, beat SD-map priors on OpenLane-V2 even under a new split.","key_machinery":"The load-bearing object is the trajectory prior dataset built by aligning Argoverse2 motion-forecasting traces to OpenLane-V2 frames using vehicle poses, filtered by length, smoothed, and pruned to frames where trajectory counts exceed five times the centerline count. The paper feeds this prior into the network in two encodings: a rasterized heatmap that stores, per grid cell, the normalized visit density and average movement direction, and vectorized instance tokens obtained by k-means clustering or farthest-point sampling with Frechet distance. The fusion is done either by element-wise addition of the heatmap into BEV features during encoding or by initializing extra decoder queries and reference points from the vectorized tokens. A dedicated alignment module then concatenates prior and BEV features, predicts 2D offsets for each position, warps the prior with bilinear interpolation, and combines the two feature maps through softmax-normalized confidence weights, with a BEV segmentation loss providing supervision for the learned alignment.","core_discovery":"The core discovery is that the path history of ordinary vehicles, treated as a map prior, captures the geometric structure of lanes more faithfully than road-level SD maps, especially at intersections, and that this fidelity translates into measurable gains in online mapping. The paper quantifies the geometric advantage by a centerline-aligned IoU of 0.39 for trajectories versus 0.16 for SD maps. It then shows, through systematic ablations, that the trajectory prior improves performance no matter how it is encoded or fused; the best result comes from adding the heatmap to BEV features and combining it with an SD-map prior, reaching 42.30 APls and 34.65 T OPlsls on the validation split, and 22.50 APls and 21.22 T OPlsls on the StreamMapNet split. The authors present this as evidence that trajectory priors are a better source of supplemental lane information than existing explicit map priors.","pith_inferences":["Beyond the paper: if trajectory priors keep their value across geographically disjoint splits, the same recipe could turn continuously collected fleet data into self-updating lane priors, reducing reliance on manual HD map annotation.","Beyond the paper: because the original split has a 54% geographic overlap between Argoverse2 train and validation, the headline +10.0/+9.25 gains likely mix in location memorization; the new-split numbers are the cleaner estimate of true prior value, and comparing them against existing methods under the same protocol would sharpen the conclusion.","Beyond the paper: the alignment module is prior-agnostic; applying it to satellite or HD-map priors, or to detect temporally shifted lane changes, is a natural transfer test.","Beyond the paper: one could measure how much of the gain comes from trajectory geometry rather than trajectory density by ablating with random synthetic traces of the same density."],"forward_implications":["Trajectory data alone outperforms SD-map prior fusion on every reported metric; combining trajectory with SD map yields the largest gains (+10.0 APls, +9.25 T OPlsls).","Heatmap-based augmentation of BEV features is the strongest single fusion scheme for topology reasoning; query initialization from vectorized tokens still improves the baseline but is slightly weaker.","The alignment module contributes a consistent gain: removing it drops APls from 42.30 to 41.13 and T OPlsls from 34.65 to 34.18.","Under the new split protocol from StreamMapNet, TrajTopo still beats the LaneSegNet baseline by a wide margin (22.50 vs 16.67 APls; 21.22 vs 17.08 T OPlsls), indicating the benefit is not solely a memorization artifact of the original split."],"supporting_citations":[{"why":"Supplies the crowdsourced motion-forecasting trajectories from which the trajectory prior is extracted and aligned.","marker":"[4]"},{"why":"Defines the OpenLane-V2 benchmark, its lane-segment and topology labels, and the evaluation protocol used for all comparisons.","marker":"[13]"},{"why":"Provides the LaneSegNet baseline architecture and loss formulation that the paper extends with trajectory and SD-map priors.","marker":"[9]"},{"why":"Establishes the SD-map prior fusion framework and the SMERF baseline that trajectory priors are compared against.","marker":"[1]"},{"why":"Supplies the new split protocol and highlights the geographic overlap issue motivating the cross-scenario evaluation.","marker":"[18]"},{"why":"Contributes the LaneDAG topology head adopted for lane-segment topology reasoning.","marker":"[25]"},{"why":"Offers the BEVFormer encoder used to build bird's-eye-view features that the trajectory heatmap is fused into.","marker":"[29]"}],"fun_headline_variants":["Trajectory priors boost lane perception and topology reasoning","Crowdsourced driving traces enhance lane map accuracy","Trajectory priors outperform SD maps for lane segments","Argoverse2 trajectories improve lane topology by 9.25","Lane perception gains from driver path history priors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that motion-forecasting trajectories gathered from the same cities at different times and under different scenarios can be pose-aligned to perception frames and used as honest lane-structure priors, rather than as a way of leaking validation geography into the model.","fun_headline_variants_meta":{"raw":{"variants":["Trajectory priors boost lane perception and topology reasoning","Crowdsourced driving traces enhance lane map accuracy","Trajectory priors outperform SD maps for lane segments","Argoverse2 trajectories improve lane topology by 9.25","Lane perception gains from driver path history priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000746,"raw_usage":{"total_tokens":3330,"prompt_tokens":958,"completion_tokens":2372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":2305}},"tokens_in":574,"tokens_out":2372,"duration_ms":16584,"temperature":1.0,"reasoning_tokens":2305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:26:16.718593+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full comparison on a split where the trajectory-collection and evaluation areas are strictly disjoint (for example, trajectories from one set of cities and perception frames from another); if the trajectory prior then gives no consistent improvement over the SD-map baseline, the reported gains are explained by geographic overlap rather than by the lane-structure content of the priors.","supporting_citations":[{"cited_title":"Openlane-v2: A topology reasoning bench- mark for unified 3d hd mapping,","cited_arxiv_id":null,"evidence_quote":"Defines the OpenLane-V2 benchmark, its lane-segment and topology labels, and the evaluation protocol used for all comparisons."},{"cited_title":"Augmenting lane perception and topol- ogy understanding with standard definition navigation maps,","cited_arxiv_id":null,"evidence_quote":"Establishes the SD-map prior fusion framework and the SMERF baseline that trajectory priors are compared against."},{"cited_title":"Streammapnet: Streaming mapping network for vectorized online hd map construction,","cited_arxiv_id":null,"evidence_quote":"Supplies the new split protocol and highlights the geographic overlap issue motivating the cross-scenario evaluation."},{"cited_title":"Lanedag: Automatic hd map topology generator based on geometry and attention fusion mechanism,","cited_arxiv_id":null,"evidence_quote":"Contributes the LaneDAG topology head adopted for lane-segment topology reasoning."}],"review_version":1}