{"id":"3d2e5bd8-df9a-4193-b737-bd036103a225","arxiv_id":"2608.05209","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding bidirectional vector and raster consistency losses to HD map training improves mAP and temporal consistency across several baselines on nuScenes and Argoverse 2.","lead":"MapTCL is an extra training step that makes self-driving maps of road markings steadier from one moment to the next by comparing current and previous map predictions. It improves map accuracy and consistency on two self-driving benchmarks without slowing down the final system.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BVCL's forced Hungarian matching (Sec. III-C, Eq. 1) has no unmatched option, so high-confidence false positives or occluded/disappeared elements are always paired with some current instance, biasing the temporal consistency signal.","rationale":"The reader's weakest_assumption correctly identifies the correspondence premise. I agree it is the most load-bearing step: BVCL is the named novelty, and Eqs. 1-3 are the only mechanism that turns predictions into a training signal. The paper's own design choices (late-stage application) concede that early matches are unreliable, but no evidence is given that late-stage matches are reliable under occlusion, which is exactly the scenario targeted in Table IV. A concrete unmatched-matching experiment would isolate whether force-matching is harmful. I also note that the component ablation (Table V) shows a substantial mAP contribution from per-frame segmentation supervision, so the temporal-consistency-specific effect is smaller than the headline gain; this is a secondary attribution issue and does not change the conditional verdict. The lack of code and single runs already motivated the reader's condition; releasing code or performing the proposed oracle analysis would satisfy it.","tokens_in":13279,"tokens_out":10282,"duration_ms":122485,"concrete_test":"Retrain StreamMapNet + MapTCL on nuScenes newsplit with BTM modified to allow unmatched instances: add a rejection threshold on the Hungarian matching cost (or a dustbin row) so that pairs whose position/class cost exceeds a fixed margin are excluded from L_BVCL. Tune the margin on validation, then report mAP and C-mAP on the same newsplit validation. If forced matching is benign, the unmatched variant will not improve or preserve the +3.7/+2.8 gains; if the gains shrink or disappear, the original consistency signal relied on erroneous forced correspondences. A complementary oracle check: for each matched pair, measure whether the warped endpoints fall within 1.5 m of the same ground-truth map element in both frames.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that BVCL teaches the model useful temporal consistency. The load-bearing premise, entering at Section III-C through Eqs. 1-3 and Eq. 5, is that the bidirectional Hungarian matches identify the same physical map element. Two properties make this insecure. First, the matching cost in Eq. 1 is minimized unconditionally over all K past high-confidence instances; there is no rejection threshold or unmatched category. Any confident past prediction, including a false positive or an element that is temporarily absent from the current frame because of occlusion, is force-matched to the nearest current instance and then dragged toward the past geometry/class by L_BVCL. The confidence filter tau=0.3 is low enough to admit many false positives, and the authors explicitly defer MapTCL to late training (Section IV-A) only because early predictions are too noisy, which acknowledges the residual noise risk. Second, the backward matching is similarly forced, so the same pathology can propagate in both directions. This does not make the method impossible, but it means the reported gains could come from a biased regularizer rather than from correct correspondences; the paper offers no oracle check of matching quality.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MapTCL, an auxiliary training strategy for online vectorized HD map construction. It adds two temporal-consistency losses to a per-frame supervised baseline: Bidirectional Vector Consistency Learning (BVCL), which warps high-confidence vector instances across frames, matches them with Hungarian assignment, and penalizes geometric and semantic discrepancies; and Raster map Consistency Learning (RCL), which adds BEV segmentation supervision plus a raster-level consistency loss between current and past predicted maps. The auxiliary modules are removed at inference. Experiments on nuScenes and Argoverse 2 with StreamMapNet, SQD-MapNet, and MapTracker report consistent mAP and C-mAP gains, including +3.7 mAP and +2.8 C-mAP on the nuScenes newsplit, with unchanged FPS.","tokens_in":13560,"tokens_out":8579,"duration_ms":93068,"significance":"If the empirical claims hold, the contribution is practically useful: a training-only regularizer that improves both accuracy and temporal stability across several baselines and two benchmarks, with no additional inference overhead, would be an attractive plug-in for HD-map systems. The paper's strengths include the breadth of the experimental matrix, the long-range and occlusion analyses, and the explicit adoption of the C-mAP temporal-consistency metric. The main caveats are that (i) the C-mAP gains are partly expected by construction, since BVCL and RCL directly minimize cross-frame prediction differences; (ii) the headline numbers appear to come from single runs with hyperparameters tuned on the same validation sets used for final comparison; and (iii) the forced Hungarian matching in BVCL can pair false positives or occluded elements with unrelated instances, making the consistency signal potentially self-confirming. The mAP gains, if reproducible, would constitute a meaningful result, but the current evidence is not yet sufficient to support the strong 'consistently enhances' claim.","major_comments":[{"comment":"The matching in Eqs. (1)-(3) is forced: every selected high-confidence instance must be paired with some instance in the other frame, because there is no unmatched category or rejection threshold. Under the low confidence threshold tau=0.3, false positives and map elements that are temporarily occluded will be paired with the nearest current instance and dragged toward the past geometry/class by L_BVCL, and the backward pass does the same in reverse. The paper's own decision to apply MapTCL only in later training stages (Sec. IV-A) and the degraded results at tau=0.1 in Table VII confirm that low-confidence predictions are unreliable. This makes the central mechanism vulnerable to self-confirmation: the loss can improve by making the model agree with its own false positives. Please provide an oracle check of matching quality against ground-truth instance associations, or add a cost-threshold/unmatched option, and report results under pose noise as in the MapTracker setting.","section":"Section III-C, Eqs. (1)-(3)"},{"comment":"The key hyperparameters (tau, N_t, w_n, u_n, alpha_n, lambda_1-lambda_3, beta) are tuned by ablations on the same validation sets that are then used for the final comparisons (Tables VI and VII), and the manuscript gives no multi-seed statistics or error bars. The statement after Table II that the large C-mAP gain on the old split 'may also be affected by hyperparameter settings' is an explicit limitation of the same kind, and the Conclusion similarly notes sensitivity to temporal context length. With no held-out validation for hyperparameter selection, no multiple seeds, and no released code or checkpoints, the claim that MapTCL 'consistently enhances' baselines is not yet statistically supported. Please add at least three seeds with mean and standard deviation for the main tables, or fix hyperparameters on a separate split before evaluating the final comparisons.","section":"Section IV-A, IV-C, Tables I-III"},{"comment":"The C-mAP gains should not be presented as independent evidence of a temporal-stability benefit. C-mAP is designed to measure temporal consistency, and L_BVCL and L_RCL explicitly minimize cross-frame geometric and semantic discrepancies of predictions, so an increase in C-mAP is expected when these losses are added. The mAP improvements are the more decisive evidence that the regularizer helps beyond optimizing the evaluation metric itself. Please report an additional temporal metric not directly aligned with the training objective (for example, instance-level tracking stability or a flicker-rate measure) and clearly distinguish 'improves C-mAP by construction' from 'improves temporal consistency.'","section":"Section IV-B, Eqs. (5)-(12)"},{"comment":"The exact objective is under-specified. In Eq. (1), pi is described as a permutation of the K selected past instances, but the sum runs over current indices i=1..K while the current map has N instances; Eq. (6) uses gamma_i(j) without defining how the point-level assignment is obtained per matched instance; and Sec. III-E says the backward losses are computed by 'warping the current HD map instance to the current frame,' which appears to be a typo for warping to the past coordinate frame. These ambiguities matter because the Hungarian matching is the core of BVCL. Please rewrite the matching and loss definitions with explicit index sets, coordinate frames, and the distinction between the N predicted instances and the K selected high-confidence instances.","section":"Section III-C, III-E, Eqs. (1) and (6)"}],"minor_comments":[{"comment":"The component-ablation table is difficult to read because the check marks are not clearly aligned with the column headers in the text; please reformat so that each row's combination of VCL, BVCL, SL, and RCL is unambiguous.","section":"Table V"},{"comment":"The class matching cost L_cls is called an L1 distance between class distributions; please specify whether it is applied to probability vectors or logits and whether it is summed over classes.","section":"Section III-B and Eq. (2)"},{"comment":"The sentence 'Both datasets follow the standard 700/150/150 train/val/test split' is stated for the original splits, but the paper mainly evaluates on non-overlapping splits; please clarify the sizes of the newsplit sets as well.","section":"Section IV-A"},{"comment":"DTCLMapper is a closely related temporal-consistency method, but it appears only in the old-split tables; please add a direct comparison under the same single-frame fusion setting, or state explicitly why such a comparison is not possible.","section":"Related Work, Sec. IV-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a clear and potentially useful core idea, and the mAP gains, if robust, would be of interest to the HD-mapping community. The main technical risk is the forced matching in BVCL, because it can turn false positives into training targets; the empirical risk is the combination of single-run results, validation-set hyperparameter selection, and no code release. I would encourage the editor to treat the revision as requiring the additional experiments described in the major comments, especially the multi-seed results and the matching-quality analysis. The overlap with DTCLMapper should also be addressed explicitly in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MapTCL is a clean, useful training-time addition to vectorized HD map models. The genuinely new piece is bidirectional vector-level matching between current and past predictions; DTCLMapper, the closest prior work, uses contrastive learning and a raster occupancy loss rather than forward-backward vector alignment. The paper reports consistent gains across StreamMapNet, SQD-MapNet, and MapTracker on nuScenes and Argoverse2, with mAP improvements up to +3.7 and C-mAP up to +6.2, all at zero inference cost. That is a real contribution if it checks out.\n\nThe authors do solid empirical work: component ablations, memory size, confidence threshold, long-range and occlusion scenarios. They also acknowledge hyperparameter sensitivity, particularly the temporal context length. The methodology is clearly described.\n\nThe soft spots are real. The main one is the forced Hungarian matching in Eq. 1: every high-confidence past instance is matched to some current instance, with no unmatched or rejection option. False positives and temporarily occluded elements are therefore dragged toward the nearest current prediction, and the loss can reward agreeing with its own mistakes. The confidence threshold of 0.3 is permissive, and applying MapTCL only in later training stages is an implicit admission that early predictions are too noisy. The paper never measures matching quality, e.g., by comparing assigned pairs against ground-truth correspondences. This concern lands on reading the paper and is the thing I'd most want addressed.\n\nThe other weaknesses are methodological: single-run results without error bars, hyperparameters selected on the same validation sets used for final comparisons, and no released code or checkpoints. Also, the C-mAP metric measures exactly what BVCL optimizes, so those gains are partly by construction; the mAP gains are less directly optimized and carry the argument.\n\nThe central claim is plausible but not yet proven. I'd send this to a serious referee. The idea is solid, the experiments are extensive, and the limitations are disclosed. What's missing is independent reproduction, variance estimates, and a sanity check on the matching quality.\n\nThis paper is for researchers working on temporal consistency in online HD mapping. I'd bring it to a reading group, but I wouldn't cite it as evidence until the results are reproduced or artifacts are released.","headline":"Genuinely new bidirectional vector consistency loss for HD map training; consistent gains reported, but forced matching and single-run evidence leave the causal claim unproven.","tokens_in":14088,"tokens_out":3139,"would_cite":false,"duration_ms":31371,"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":"Adding bidirectional temporal-consistency losses between matched high-confidence map instances across frames improves both accuracy and temporal stability of online HD map construction, without adding inference cost.","keywords":["online HD map construction","temporal consistency","vectorized map","bidirectional alignment","auxiliary training loss","BEV perception","autonomous driving"],"falsifier":"Train the same model with BVCL applied to instances taken from different scenes (so the matched pairs are guaranteed not to be the same physical elements); if mAP and C-mAP gains persist, the temporal-correspondence assumption is not what drives the reported improvements.","tokens_in":13080,"feed_emoji":"🗺️","tokens_out":6706,"duration_ms":60300,"temperature":0.7,"pith_summary":"Online HD map construction from onboard cameras usually trains each frame against ground truth independently, so the same physical lane divider or road boundary can jitter, flicker, or vanish between consecutive frames. This paper proposes MapTCL, a plug-and-play auxiliary training strategy that adds two temporal consistency losses: Bidirectional Vector Consistency Learning (BVCL), which matches high-confidence vector instances from past and current frames after ego-motion warping and penalizes their geometric and class differences, and Raster map Consistency Learning (RCL), which penalizes differences between dense BEV raster maps. The claim is that this explicit cross-frame objective improves both single-frame accuracy and temporal stability for existing HD map models, without adding any inference cost. The paper reports consistent gains across three baselines on nuScenes and Argoverse2, including +3.7 mAP and +2.8 C-mAP on nuScenes for StreamMapNet. A sympathetic reader would care because the fix is training-only and can be dropped into any existing model.","feed_headline":"Training-only loss lifts HD map accuracy by up to 3.7 mAP","feed_subtitle":"Auxiliary losses that match past and current frames make online HD maps steadier, with no inference slowdown.","key_machinery":"The load-bearing mechanism is Bidirectional Temporal Matching (BTM), a two-stage matching procedure: it filters predicted vector instances to those with confidence above $\\tau$, warps them into the other frame's coordinate system using ego motion, solves instance-level assignment with the Hungarian algorithm, then solves point-level assignment within each matched pair, and repeats the process in the reverse direction. The matched pairs feed two loss terms, a SmoothL1 positional loss and a KL-divergence class loss, with temporal weights that decay for older frames. RCL is the second mechanism: a segmentation head produces a binary raster map, supervised against a rasterized ground truth, and a mask Focal loss between the current raster map and warped historical raster maps stabilizes dense BEV features. Both auxiliary losses are added to the total loss and applied only in later training stages, which the paper argues avoids propagating early false positives.","core_discovery":"On its own terms, the paper's central claim is that per-frame ground-truth supervision leaves a measurable gap: temporal fusion architectures such as StreamMapNet, SQD-MapNet, and MapTracker still lack an explicit objective that penalizes geometric noise and temporal jitter between consecutive predicted maps. MapTCL closes this gap with two auxiliary losses. BVCL selects instances with confidence above a threshold, warps them across frames using ego motion, performs Hungarian-based instance- and point-level matching in both forward and backward directions, and minimizes SmoothL1 point differences plus KL-divergence on class distributions. RCL generates a binary BEV raster map through a segmentation head and minimizes mask Focal loss between the current raster map and warped past raster maps. Trained jointly with the standard map loss, these losses are reported to improve mAP and C-mAP across all tested baselines on both benchmarks while leaving inference FPS unchanged, since the module is removed at inference time.","pith_inferences":["Editorial: the same prediction-level temporal consistency objective could be applied to other streaming perception outputs, such as online lane detection or BEV segmentation, which suffer from the same frame-to-frame jitter.","Editorial: because the confidence filter creates a curriculum that only later-stage predictions participate in, MapTCL could be combined with pseudo-labeling on unlabeled temporal sequences, an extension the paper does not explore.","Editorial: the C-mAP gains are smaller and less consistent than mAP gains in some rows of the tables; a reader should treat the temporal-stability claim as scene-dependent and in need of a dedicated temporal benchmark.","Editorial: a direct testable prediction is that BVCL's contribution shrinks as ego-motion estimates become noisier, since warping errors corrupt the matching; the paper does not quantify this sensitivity."],"forward_implications":["Existing vectorized HD map models can adopt MapTCL as a training-only plug-in, improving accuracy and temporal stability with no change to inference latency or memory footprint.","Temporal consistency can be enforced at the prediction level without recurrent feature fusion, making the benefit complementary to architectural temporal fusion and applicable to models that do not fuse features.","The confidence-filtered bidirectional matching acts as a self-supervision signal across frames, potentially reducing the need for temporally annotated ground truth.","Memory length matters: the paper's ablation shows performance peaks at $N_t=5$ and degrades at $N_t=7$, indicating that the optimal temporal context is a hyperparameter to tune.","The reported gains on two datasets and three baselines suggest the improvement transfers across model designs and perception ranges."],"supporting_citations":[{"why":"provides the StreamMapNet baseline that MapTCL is first integrated into and the two-stage training schedule.","marker":"[13]"},{"why":"supplies the hierarchical instance- and point-level matching strategy used in Bidirectional Temporal Matching.","marker":"[9]"},{"why":"supplies the BEV segmentation head and raster ground-truth supervision that RCL builds on.","marker":"[11]"},{"why":"provides the cycle-consistency idea from scene flow estimation that motivates bidirectional alignment.","marker":"[25]"},{"why":"the Hungarian algorithm solves the optimal instance- and point-level assignments in BTM.","marker":"[26]"},{"why":"provides the C-mAP temporal consistency metric and serves as a stronger baseline (MapTracker) that MapTCL also improves.","marker":"[15]"},{"why":"one of the two evaluation benchmarks (nuScenes) used for all experiments.","marker":"[17]"},{"why":"the second evaluation benchmark (Argoverse2) used for all experiments.","marker":"[18]"}],"fun_headline_variants":["Training-only losses add 3.7 mAP to HD map builders","Auxiliary losses steady HD maps: +3.7 mAP, no slowdown","Bidirectional alignment loss boosts HD map accuracy","Plug-and-play training loss lifts HD map mAP by 3.7","Temporal consistency loss improves HD maps at no cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"High-confidence predicted instances in past and current frames, after ego-motion warping and Hungarian matching, actually correspond to the same physical map elements; if warping is inaccurate or false positives pass the confidence filter, the consistency loss rewards the model for agreeing with its own mistakes.","fun_headline_variants_meta":{"raw":{"variants":["Training-only losses add 3.7 mAP to HD map builders","Auxiliary losses steady HD maps: +3.7 mAP, no slowdown","Bidirectional alignment loss boosts HD map accuracy","Plug-and-play training loss lifts HD map mAP by 3.7","Temporal consistency loss improves HD maps at no cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000325,"raw_usage":{"total_tokens":1831,"prompt_tokens":961,"completion_tokens":870,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":780}},"tokens_in":577,"tokens_out":870,"duration_ms":8430,"temperature":1.0,"reasoning_tokens":780,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:06:38.387032+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same model with BVCL applied to instances taken from different scenes (so the matched pairs are guaranteed not to be the same physical elements); if mAP and C-mAP gains persist, the temporal-correspondence assumption is not what drives the reported improvements.","supporting_citations":[{"cited_title":"Streammapnet: Streaming mapping network for vectorized online hd map construc- tion,","cited_arxiv_id":null,"evidence_quote":"provides the StreamMapNet baseline that MapTCL is first integrated into and the two-stage training schedule."},{"cited_title":"Maptrv2: An end-to-end framework for online vectorized hd map construction,","cited_arxiv_id":null,"evidence_quote":"supplies the BEV segmentation head and raster ground-truth supervision that RCL builds on."},{"cited_title":"Just go with the flow: Self- supervised scene flow estimation,","cited_arxiv_id":null,"evidence_quote":"provides the cycle-consistency idea from scene flow estimation that motivates bidirectional alignment."},{"cited_title":"Maptracker: Track- ing with strided memory fusion for consistent vector hd mapping,","cited_arxiv_id":null,"evidence_quote":"provides the C-mAP temporal consistency metric and serves as a stronger baseline (MapTracker) that MapTCL also improves."}],"review_version":1}