{"id":"5e46ec2a-685f-446f-8a6b-08f2e9d8f6b8","arxiv_id":"2608.12600","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A teacher-student framework with Beta-confidence maps and spatial clipping produces pseudo-labels that improve online HD mapping by +6.1 mAP using only 16.5% labeled data.","lead":"The authors propose a semi-supervised pipeline for online HD map construction that generates training labels from unlabeled driving data by keeping only high-confidence segments of predicted map elements. The method reports a +6.1 mAP gain over a labeled-only baseline on nuScenes, which could cut the cost of annotating high-definition maps for autonomous driving.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation-set temperature calibration leaks into pseudo-label generation, so the reported +6.1 mAP gain is not a clean generalization estimate.","rationale":"The reader's weakest_assumption emphasizes the teacher-signal assumption and notes the temperature-calibration issue as secondary. I agree the teacher-signal assumption is acknowledged and at least partially supported by Table 1, where pseudo-label quality improves with clipping. However, the most load-bearing concern for the central claim is the validation-set contamination through temperature calibration: the paper explicitly states T* is optimized on the nuScenes validation set, and the same validation set is used for the headline Table 2 evaluation. Because T* affects the Beta confidence map and clipping, it changes the pseudo-labels that train the student; therefore the final mAP is not a clean out-of-sample measurement. The paper's ranking-based defense is incomplete because the downstream pipeline is not rank-invariant in T. This does not necessarily doom the method—the effect size might survive a clean calibration split—which is why the appropriate verdict remains conditional rather than rejection. The lack of error bars is real but secondary; a +6.1 mAP gap is large, and the more urgent question is whether the evaluation protocol is valid. Had the paper used a separate calibration set or reported sensitivity to T, this concern would be substantially mitigated. I therefore keep the reader's CONDITIONAL verdict unchanged, but with a sharper, more specific condition: re-evaluate with a clean calibration protocol.","tokens_in":11414,"tokens_out":7248,"duration_ms":79846,"concrete_test":"Re-run the full pipeline with temperature fixed at T=1 (no calibration) and with T chosen on a held-out calibration split drawn from the labeled training set (e.g., a 10% subset of DL), then recompute Table 2. If Ours (Clipping) retains a large gain (>~5 mAP) over Baseline and still beats Filtering by >~2 mAP, the validation-set leakage is not the driver. Also report a sensitivity sweep of Table 2 over T ∈ {0.5, 0.7, 1.0, 1.5} with p fixed at 30; if the headline mAP swing exceeds ~2 points, the reported single-number gain is unstable under the contaminated hyperparameter.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 ('Probability Calibration') optimizes the temperature T* on the nuScenes validation set using Eq. (3). That T* enters Eq. (4) to build α(x,y,c), which determines the Beta-distribution confidence map, spatial clipping thresholds, and ultimately the pseudo-labels used to train the student. The final mAP in Table 2 is then measured on the same validation set. Therefore, ground-truth labels from the evaluation set have influenced the pseudo-labels that produced the evaluated student. The paper's defense—that temperature scaling preserves ranking, so mAP fairness is not compromised—only covers the teacher's per-element confidence ordering for a fixed teacher. It does not address the downstream dependence: the aggregate Beta-map confidences are sums of temperature-scaled scores over varying numbers of observations, and these sums can reorder across T, changing the percentile thresholds and the clipped pseudo-labels. Even if the final mAP ranking were unaffected for a fixed teacher, the student's training signal is not fixed; it is a function of T. Thus the +6.1 mAP improvement over baseline is at least partially a tuned-on-test result rather than an unbiased SSL generalization measurement. A related protocol concern is the selection of percentile p=30 using ground-truth labels on the pseudo-unlabeled set (Table 1); while that uses training-split labels rather than the validation split, it still violates the stated unlabeled-data assumption. The temperature leakage is more direct because it uses the exact evaluation set for a decision that changes the training data of the evaluated model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PseudoMapLabeler (PML), a teacher-student semi-supervised learning framework for online vectorized HD mapping. A Uni-PrevPredMap (UPPM) teacher is trained on only 16.5% of the StreamMapNet geosplit training set. Its predictions on the remaining 83.5% are accumulated in scene-local coordinates, converted into per-class Beta-distribution confidence maps using temperature-scaled scores, and spatially clipped via percentile thresholds to produce refined map priors. The teacher is then re-applied with these priors to generate pseudo-labels, which are used to train a student UPPM or MapTR from scratch before fine-tuning on the labeled subset. On the nuScenes validation set, the paper reports +6.1 mAP over a temporal-prior-enabled UPPM baseline and +5.0 mAP for a MapTR student. The paper also ablates the percentile threshold and compares spatial clipping with element-wise filtering.","tokens_in":11670,"tokens_out":6327,"duration_ms":61849,"significance":"If the reported gains hold under a clean evaluation protocol, PML would be a useful contribution to an underexplored problem: semi-supervised learning for online vectorized HD mapping. The paper takes several steps in the right direction: it uses the geographically disjoint StreamMapNet split to prevent spatial leakage, it compares against a strong temporal-prior-enabled UPPM baseline rather than a naive single-frame model, it includes a second architecture (MapTR) to support model-agnosticism, and it makes the pseudo-unlabeled set and its ground truth available for diagnostic evaluation. The spatial-clipping idea, which preserves partial high-confidence segments instead of discarding whole elements, is reasonable and potentially valuable. However, the current evidence is weakened by two protocol issues in the experimental design: the temperature used in pseudo-label generation is calibrated on the same validation set used for final evaluation, and the percentile threshold is selected using ground truth labels from the pseudo-unlabeled set. In addition, the results are presented without variance estimates.","major_comments":[{"comment":"The temperature parameter T* is optimized on the nuScenes validation set via Eq. (3) and then used in Eq. (4) through the calibrated scores scalib, which determine the Beta confidence maps, the percentile thresholds of Eq. (8), and ultimately the pseudo-labels that train the student. The final mAP in Table 2 is measured on the same validation set. The paper's argument that temperature scaling preserves per-element confidence ranking does not cover the downstream aggregation: the sums in Eq. (4) combine varying numbers of observations per grid cell and can reorder across T, and the percentile thresholds in Eq. (8) shift with T. The student's training signal is therefore a function of validation-set labels, making the reported +6.1 mAP a tuned-on-validation result rather than an unbiased estimate. I ask the authors to calibrate T on a split disjoint from the final evaluation set (for example, a held-out portion of the labeled training data or a separate split of the pseudo-unlabeled set) and re-report Table 2, or to demonstrate explicitly that the results are insensitive to T over a plausible range.","section":"Section 4.2, Eq. (3), Eq. (4), Eq. (8), Table 2"},{"comment":"The percentile p=30 used for the final student is selected by evaluating pseudo-label quality on the pseudo-unlabeled set using its ground truth labels (Table 1). In a genuine SSL scenario, the unlabeled set would not have labels; using those labels to choose p means the training signal is selected with information that the method would not have access to in deployment. This is distinct from evaluating a fixed pipeline on a held-out set. The current protocol also creates an internal tension, since mAP peaks at p=20 while Dice/IoU peaks at p=30 and the justification is post hoc. I recommend selecting p without DU labels (e.g., by cross-validation on DL or by a fixed confidence heuristic) and reporting how much Table 2 changes across p=20, p=30, and p=40.","section":"Section 4.2, Table 1"},{"comment":"All results in Table 2 appear to be from single training runs; no standard deviations, number of seeds, or significance tests are reported. The headline improvements (+6.1 mAP over the UPPM baseline, +2.8 mAP over element filtering, and +5.0 mAP for MapTR) are moderate in size and could be within run-to-run variation for models of this complexity. Please report mean and standard deviation over at least three runs (or otherwise establish reproducibility), since the central empirical claim depends on these differences being reliable.","section":"Table 2"}],"minor_comments":[{"comment":"The displayed formula for the accumulated map elements is garbled in the manuscript ('M acc = Snf rm(scene) i=1 Mscene i'); please define n_frames(scene) explicitly and write the union operation cleanly.","section":"Section 3.4"},{"comment":"The text says the teacher is initially trained without any map priors, yet the baseline is described as having temporal priors enabled. Please clarify the distinction between the model's internal temporal priors (its own past predictions) and the proposed refined map priors, so the reader can see exactly what is being ablated.","section":"Section 3.7"},{"comment":"The GT Prior row reports values in parentheses without explanation; please add a footnote stating whether these are computed with ground-truth priors during the second pass and why they are parenthesized.","section":"Table 1"},{"comment":"The statement that the pseudo-unlabeled set contains ground truth labels 'used only for evaluation purposes' is inconsistent with the use of those labels to select p in Table 1; please rephrase once the protocol is fixed.","section":"Section 4.2"},{"comment":"The term 'map priors' is used both for the proposed refined priors and for the prior-fusion mechanisms in related work (NMP, P-MapNet); consider using a distinct term such as 'refined pseudo-priors' to reduce ambiguity.","section":"Throughout"},{"comment":"There are minor typographical inconsistencies in the equations, such as missing boldface for vectors in Eq. (1) and inconsistent placement of subscripts and superscripts; a careful proofread would improve readability.","section":"Equations"}],"recommendation":"major_revision","confidential_remarks":"The main gate for this paper is the evaluation protocol. If the authors can show that the results are unchanged when T is calibrated on a split disjoint from the final evaluation set and when p is selected without using labels from the pseudo-unlabeled set, I would see no further obstacle to publication. The paper is otherwise clearly written, the baseline is appropriately strong, and the cross-architecture experiment is a nice touch."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper: it has a genuinely new pseudo-label refinement idea—segment-level spatial clipping driven by a Beta-distribution confidence map—and a careful experimental setup. But the headline +6.1 mAP is not a clean generalization estimate because the temperature used to build pseudo-labels is tuned on the validation set that the final mAP is measured on.\n\nWhat's genuinely new: the spatial clipping technique preserves high-confidence sub-segments of predicted polylines instead of discarding whole elements, and the Beta map encodes both observation count and per-detection confidence. The experimental design is also serious: geographically disjoint split, a temporal-prior-enabled baseline, and a cross-architecture check with MapTR. The direct comparison against element-level filtering is the cleanest evidence for the technique; +2.8 mAP over filtering is a solid within-paper result.\n\nThe soft spot is the evaluation protocol. The stress-test concern is valid. Temperature scaling preserves per-element ranking, but the Beta map aggregates sums of scaled scores over different numbers of observations, so cell confidence orderings—and hence the percentile thresholds and the clipped pseudo-labels—can change with T. The student's training signal is therefore a function of a parameter chosen using the very labels on which the final mAP is reported. That makes +6.1 mAP a partially tuned-on-test number. It doesn't invalidate the spatial-clipping idea, but the unbiased gain is likely smaller. The selection of p=30 using ground truth on the pseudo-unlabeled set is a related problem, even though it's on the training split. And there are no error bars anywhere; every result is a single run.\n\nThis paper is for people working on SSL for online HD mapping. The spatial clipping contribution is worth knowing, and the weaknesses are fixable. I'd send it to peer review with a demand for a separate calibration split, a sensitivity analysis over T, and multiple seeds. As is, the method deserves attention; the specific numbers should not be taken at face value.","headline":"A clever segment-level pseudo-label refinement idea, but the headline mAP gain is compromised by temperature tuning on the evaluation set.","tokens_in":12229,"tokens_out":3364,"would_cite":true,"duration_ms":33942,"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":"The paper claims that a confidence-aware spatial-clipping pipeline can turn noisy teacher predictions on unlabeled driving data into pseudo-labels strong enough to lift semi-supervised online HD map construction by 6.1 mAP.","keywords":["semi-supervised learning","online HD map construction","pseudo-labels","confidence calibration","spatial clipping","Beta-distribution confidence map","teacher-student framework","nuScenes"],"falsifier":"Compare the precision of clipped high-confidence segments with the discarded low-confidence segments on the pseudo-unlabeled set, whose labels are available for evaluation; if top-percentile segments are not substantially more accurate than the discarded tail, the confidence map is not separating signal from noise and the 6.1 mAP gain must come from elsewhere, such as the extra pretraining data regardless of quality. A targeted control would also corrupt ego poses during temporal accumulation and check whether the mAP gain collapses, which would reveal dependence on localization accuracy.","tokens_in":11217,"feed_emoji":"🗺️","tokens_out":6405,"duration_ms":52577,"temperature":0.7,"pith_summary":"The paper claims that a semi-supervised teacher–student pipeline can generate usable pseudo-labels for online high-definition (HD) map construction from a dataset that is only 16.5% labeled, without discarding too much information. Instead of deleting entire predicted map elements that look unreliable, it keeps high-confidence pieces: predictions are accumulated across frames, a per-cell confidence map is built with a Beta distribution, and polylines are clipped to the confident regions. Those clipped pieces are fed back to the teacher as map priors, which sharpens its second pass over unlabeled data; the resulting pseudo-labels pre-train a student that is then fine-tuned on the labeled set. On the nuScenes StreamMapNet geosplit this adds 6.1 mean average precision (mAP) over a comparable labeled-only baseline, and the same pseudo-labels give a different architecture a 5.0 mAP gain, supporting the claim that the pipeline is architecture-agnostic.","feed_headline":"Confidence clipping lifts semi-supervised map accuracy by 6.1 mAP","feed_subtitle":"A teacher–student pipeline turns unlabeled driving video into precise map priors for online HD map construction.","key_machinery":"The load-bearing object is the Beta-distribution confidence map together with percentile-based spatial clipping. For each 0.5 m BEV cell and semantic class, the map updates Beta shape parameters $\\alpha$ and $\\beta$ from the number of times the cell was observed and the sum of calibrated prediction confidences, so a cell's posterior mean confidence, $\\text{conf}(x,y,c) = \\alpha/(\\alpha+\\beta)$, rises with repeated high-confidence detections and is damped by a weak prior biased toward absence ($p_0=0.2$, $\\kappa=2.0$). Confidence values are sampled along accumulated polylines, and only contiguous runs above the class threshold $\\tau_c = \\operatorname{Percentile}(\\{c_i\\}, 100-p)$ survive, with polygons kept only if all points pass. This object carries the argument by deciding which partial geometry is trustworthy enough to re-inject as a prior into the teacher's prior-conditioned decoder.","core_discovery":"On its own terms, the paper's discovery is that confidence-aware spatial clipping turns noisy temporally accumulated predictions into map priors good enough to bootstrap a better pseudo-label generator. The teacher is first trained only on the labeled subset; its per-frame predictions are calibrated by temperature scaling, accumulated in a scene-local frame, and summarized by a Beta-distribution confidence map per 0.5 m bird's-eye-view cell. Rather than thresholding whole elements, the method clips each polyline to continuous runs of points whose confidence exceeds a class-specific percentile threshold, and uses the result as a prior for a second teacher pass over the unlabeled data. Those refined predictions, used as pseudo-labels, pre-train a student from scratch before fine-tuning on the labeled subset. The paper reports +6.1 mAP over the labeled-only baseline and +2.8 mAP over whole-element filtering, with the largest per-class gain on pedestrian crossings.","pith_inferences":["If the gain is driven by confidence separation rather than simply more pretraining data, the same pipeline should transfer to lane-detection or BEV segmentation outputs that can be rasterized onto the same confidence grid.","Because the method depends on accurate ego poses during temporal accumulation, localization noise is a hidden hyperparameter; a controlled pose-jitter study would show whether the 6.1 mAP gain degrades faster than single-frame baselines.","The gap to the ground-truth-prior ceiling points to teacher quality as the next bottleneck, so an iterative self-training loop that retrains the teacher on its own high-confidence pseudo-labels is the natural next step."],"forward_implications":["With 16.5% labeled data, a student trained on pseudo-labels from spatial clipping and fine-tuned on labeled data reaches 27.6 mAP on the nuScenes validation set versus 21.5 mAP for the labeled-only baseline.","Spatial clipping beats element-level filtering by 2.8 mAP for UPPM and 1.4 mAP for MapTR, showing that retaining partial reliable segments is better than all-or-nothing element decisions.","The same pseudo-labeled dataset transfers to a different architecture, giving a MapTR student a 5.0 mAP gain, which supports the claim that the refinement-and-pseudo-labeling pipeline is model-agnostic.","Pseudo-label quality peaks around percentile p=20–30 on the pseudo-unlabeled set, with rasterized Dice/IoU favoring p=30 while the mAP difference from p=20 is within noise.","The roughly 10 mAP gap between refined pseudo-labels and the ground-truth-prior upper bound marks headroom attributable to limited teacher quality at 16.5% labeled data."],"supporting_citations":[{"why":"Uni-PrevPredMap supplies the teacher and student architecture whose prior-conditioning mechanism is repurposed to accept refined map priors instead of ground-truth maps.","marker":"[23]"},{"why":"PrevPredMap provides the temporal previous-prediction prior that the teacher builds on, and the paper's prior refinement inherits this mechanism.","marker":"[24]"},{"why":"MapTR is the second architecture used to test cross-architecture transfer of the pseudo-labels, producing the +5.0 mAP result.","marker":"[14]"},{"why":"StreamMapNet defines the geographically disjoint train/validation split that the low-label experiment uses to avoid spatial leakage.","marker":"[33]"},{"why":"Temperature scaling from this work calibrates the teacher's confidence scores before they are accumulated into Beta-distribution confidence maps.","marker":"[7]"},{"why":"nuScenes is the dataset for all experiments, including the 16.5% labeled / 83.5% pseudo-unlabeled split.","marker":"[2]"},{"why":"This analysis of data leakage in online mapping datasets motivates the geographically disjoint evaluation used for the claimed mAP gains.","marker":"[16]"}],"fun_headline_variants":["Spatial clipping boosts semi-supervised map accuracy by 6.1 mAP","Teacher-student with spatial clipping adds 6.1 mAP for HD maps","Confidence-aware clipping: +6.1 mAP for semi-supervised online mapping","Semi-supervised HD maps gain 6.1 mAP via spatial clipping","PseudoMapLabeler: smarter pseudo-labels gain 6.1 mAP on HD maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the teacher's predictions on unlabeled data, noisy as they are from only 16.5% labeled training, contain enough reliable signal that confidence-based clipping keeps the correct segments and trims the wrong ones; if the teacher is confidently wrong, the clipping step preserves those errors as pseudo-labels.","fun_headline_variants_meta":{"raw":{"variants":["Spatial clipping boosts semi-supervised map accuracy by 6.1 mAP","Teacher-student with spatial clipping adds 6.1 mAP for HD maps","Confidence-aware clipping: +6.1 mAP for semi-supervised online mapping","Semi-supervised HD maps gain 6.1 mAP via spatial clipping","PseudoMapLabeler: smarter pseudo-labels gain 6.1 mAP on HD maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3441,"prompt_tokens":944,"completion_tokens":2497,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":2385}},"tokens_in":560,"tokens_out":2497,"duration_ms":18256,"temperature":1.0,"reasoning_tokens":2385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:03:51.189545+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the precision of clipped high-confidence segments with the discarded low-confidence segments on the pseudo-unlabeled set, whose labels are available for evaluation; if top-percentile segments are not substantially more accurate than the discarded tail, the confidence map is not separating signal from noise and the 6.1 mAP gain must come from elsewhere, such as the extra pretraining data regardless of quality. A targeted control would also corrupt ego poses during temporal accumulation and check whether the mAP gain collapses, which would reveal dependence on localization accuracy.","supporting_citations":[{"cited_title":"In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)","cited_arxiv_id":null,"evidence_quote":"PrevPredMap provides the temporal previous-prediction prior that the teacher builds on, and the paper's prior refinement inherits this mechanism."},{"cited_title":"In: International Conference on Learning Representations (ICLR) (2023) 3, 13, 14","cited_arxiv_id":null,"evidence_quote":"MapTR is the second architecture used to test cross-architecture transfer of the pseudo-labels, producing the +5.0 mAP result."},{"cited_title":"In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)","cited_arxiv_id":null,"evidence_quote":"StreamMapNet defines the geographically disjoint train/validation split that the low-label experiment uses to avoid spatial leakage."},{"cited_title":"In: International conference on machine learning","cited_arxiv_id":null,"evidence_quote":"Temperature scaling from this work calibrates the teacher's confidence scores before they are accumulated into Beta-distribution confidence maps."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"nuScenes is the dataset for all experiments, including the 16.5% labeled / 83.5% pseudo-unlabeled split."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"This analysis of data leakage in online mapping datasets motivates the geographically disjoint evaluation used for the claimed mAP gains."}],"review_version":1}