acceptodds
Withdrawn from ICLR 2027

DUMap: Dirichlet Uncertainty-Guided Online Vectorized HD Mapping under Environmental and Geographic Domain Shifts

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Online vectorized HD mapping is important for autonomous driving, but exist- ing methods suffer severe performance degradation under environmental domain shifts such as nighttime, rain, and low visibility. Existing methods mainly model uncertainty in geometric outputs, while the evidence-allocation process remains deterministic. Under distribution shifts, a map query can also be uncertain about which sampled BEV evidence should be used. To address this issue, we propose DUMap to model query-to-BEV evidence-allocation uncertainty before geometric prediction. Dirichlet Attention (DA) models the normalized evidence-allocation weights with a Dirichlet distribution. Its mean determines the relative evidence allocation, and its total concentration measures allocation reliability. Uncertainty- Guided Residual Fusion (UGRF) uses this reliability to combine local geometric predictions with global structural refinement. We also introduce Env-Split on nuScenes by excluding nighttime and rainy samples from training. Experiments show that DUMap achieves state-of-the-art (SOTA) performance under environ- mental OOD evaluation while maintaining strong results on region-based and city-based splits.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
HD map generation, Auto driving perception

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UWbHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents DUMap, an uncertainty-aware framework for online vectorized HD map construction under environmental and geographic domain shifts. The proposed Dirichlet Evidence Allocation (DEA) module models query-to-BEV attention weights using a Dirichlet distribution, where the concentration parameter is used to estimate evidence-allocation uncertainty. The Uncertainty-Guided Residual Fusion (UGRF) module then uses this uncertainty to adaptively combine local geometric predictions with global structural refinement. Additionally, the authors introduce a low-frequency guidance branch and an environment-based evaluation split (Env-Split) on nuScenes.

Experiments on nuScenes and Argoverse 2 show consistent improvements over several existing mapping methods under environmental, region-based, and city-based evaluation protocols.

Critical strengths
  1. Relevant problem formulation and consistent empirical improvements. The paper addresses the important challenge of generalizing online vectorized mapping models to adverse environmental conditions and unseen geographic regions. The experimental evaluation covers two datasets and multiple domain-shift protocols. DUMap achieves consistent performance improvements over the evaluated baselines, suggesting that the proposed framework can improve mapping robustness.
  2. An interesting attempt to model uncertainty at the evidence-allocation stage. Rather than exclusively modeling uncertainty in the final geometric predictions, the paper introduces uncertainty modeling into BEV feature aggregation. The combination of Dirichlet attention and uncertainty-guided structural refinement provides an interesting direction for uncertainty-aware perception. The component ablations and fixed-weight fusion comparisons offer initial evidence for the effectiveness of the proposed design.
Critical weaknesses
  1. Insufficient justification that Dirichlet concentration reflects meaningful evidence uncertainty. The central methodological assumption is that the total Dirichlet concentration provides a reliable estimate of query-to-BEV evidence uncertainty under distribution shifts. However, this connection is not sufficiently established. Theorem 1 characterizes the variance induced by stochastic Dirichlet attention under fixed sampled features and fixed expected attention weights. Although the derivation is mathematically reasonable, it does not establish that the resulting concentration-based uncertainty corresponds to actual observation unreliability or geometric prediction error under environmental shifts. Moreover, the concentration parameters are learned indirectly through the mapping objective, without an explicit mechanism ensuring that concentration tracks evidence reliability. The reported uncertainty-error correlations provide some empirical support, but they do not fully resolve this concern. Consequently, the theoretical analysis and empirical evidence do not yet sufficiently justify the central interpretation of the Dirichlet concentration as a reliable OOD uncertainty signal.

  2. Limited isolation of the proposed uncertainty mechanism and potential confounding factors in environmental evaluation. Although Tables 3 and 4 provide useful component ablations and comparisons against fixed fusion weights, the experiments do not fully isolate whether the improvements arise specifically from Dirichlet uncertainty estimation rather than additional feature processing, model capacity, stochastic regularization, or adaptive fusion. In particular, stronger capacity-matched comparisons against deterministic attention and alternative attention-derived confidence measures would help establish the necessity of the Dirichlet formulation. This is especially important given the conceptual similarities to existing uncertainty-aware mapping methods such as UIGenMap. Furthermore, Env-Split is constructed using scene-description keywords. While useful as an initial evaluation protocol, it does not strictly isolate illumination and weather changes from other factors. The supplementary statistics show differences in pedestrian density between training and validation scenes, and the paper does not sufficiently control for geographic overlap or other scene-level confounders in this split. Therefore, the reported improvements are encouraging, but the evidence is not yet sufficient to establish the proposed mechanism as the principal explanation for the gains under environmental domain shifts.

Questions for authors

Q1. Validation of evidence-allocation uncertainty. How do the authors verify that the learned Dirichlet concentration reflects actual visual evidence reliability rather than merely the scale of the learned attention parameters? In particular, how does DEA uncertainty change between clear and adverse conditions, and does its relationship with geometric error remain consistent across different environment categories? A controlled analysis of concentration and prediction error under different environmental conditions would materially strengthen the central claim. Q2. Necessity of Dirichlet modeling. Can the authors provide a capacity-matched comparison with deterministic attention and alternative confidence signals, such as attention entropy, while retaining the same low-frequency branch and local-global fusion architecture? Such a comparison would clarify whether the improvement is specifically attributable to the proposed Dirichlet uncertainty estimation. Q3. Reliability of the Env-Split evaluation. How do the authors control for geographic overlap and scene-distribution differences between the favorable-condition training set and adverse-condition validation set? Would the reported improvements persist under geographically controlled environmental comparisons or paired observations of the same locations under different conditions? This clarification is important for attributing the reported performance gains to robustness against environmental shifts.

Reviewer_fVjcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses online vectorized HD mapping under changes in weather, illumination, and geographic location. DUMap models map-query allocation over sampled BEV features using a Dirichlet distribution. Its total concentration provides a query-level uncertainty signal that controls fusion between local geometric predictions and global structural refinement. The method also introduces a low-frequency auxiliary branch. Evaluation covers environmental, city-based, and region-based splits on nuScenes, together with geographic splits on Argoverse 2.

Critical strengths

1、The performance improvements are consistent across datasets and evaluation protocols. With a ResNet-50 backbone and 24 training epochs, DUMap improves over the strongest evaluated baseline by 2.8, 2.4, and 1.9 mAP points on the three nuScenes splits. It also improves by 1.4 and 2.3 points on the Argoverse 2 region-based and city-based splits (Tables 1 and 2). The gains extend beyond Env-Split to the geographic evaluations on both datasets, giving a broader picture of the method’s performance under distribution shifts.

2、Adding DEA increases mAP from 32.6 to 35.3 over the uncertainty-aware baseline, and the complete model reaches 39.2. Adaptive fusion improves by 1.1 points over the best tested fixed weight. Divider AP increases from 34.2 to 37.4 and boundary AP from 45.2 to 47.7, while pedestrian-crossing AP decreases from 34.8 to 32.6. The overall improvement therefore comes with a trade-off across map classes. The appendix also reports class-specific recall under adverse conditions. For nighttime pedestrian crossings, recall increases from 50.32% to 58.40% relative to the uncertainty-aware baseline, with false negatives decreasing from 695 to 582 (Tables 3, 4, and 14).

3、DEA separates the expected allocation from its concentration, allowing queries with similar mean attention patterns to have different allocation dispersions. It addresses a different source of uncertainty from the Laplace coordinate head: DEA concerns the allocation of BEV features, while the Laplace head concerns the predicted coordinates.

Critical weaknesses

1、The validation of DEA as a reliability signal is weaker than the validation of overall mapping performance. DEA has Pearson and Spearman correlations of 0.3943 and 0.3828 with geometric error. Table 5 also compares DEA with a Gaussian attention alternative under the same fixed fusion weight of ω = 0.5. DEA has higher uncertainty–error correlations and reaches 38.1 mAP versus 37.2 in this comparison. The failure-detection and selective-risk results in the appendix, including AUROC 0.8797 and AUSE 0.0295, concern Laplace coordinate uncertainty rather than DEA. The theorem characterizes stochastic aggregation variance under fixed sampled features and expected allocation. The paper explicitly notes that cross-query variability also depends on disagreement among the sampled features. For the fusion rule, the relevant empirical question is whether queries with higher DEA uncertainty benefit more from global refinement. A gate predicted directly from query features, using the same refinement path and training budget, would provide a baseline for assessing the contribution of the concentration signal.

2、Env-Split does not sufficiently isolate environmental factors from other scene differences. The split uses keywords such as “rain,” “night,” “evening,” and “wet” in scene descriptions. The appendix reports scene deduplication, traffic counts, map-element statistics, and the composition of adverse conditions. The geographical distribution of the two partitions and the accuracy of the keyword-derived condition labels are not reported. Pedestrian counts also differ substantially between partitions, falling from 7.68 to 3.73 per sample. Vehicle counts and the total number of map elements are much closer between the two partitions. The split is useful for evaluating adverse-condition performance. Reporting the scene distribution by map location would clarify how much geographic change accompanies the change in weather and illumination.

3、Results from repeated training runs would help assess the stability of the smaller ablation gains, particularly the 1.1-point improvement from adaptive fusion. Please report the means and standard deviations for the fixed-weight and adaptive variants under the same training settings.

Questions for authors

1、When predictions are grouped by DEA uncertainty, how do the errors of the local path and the global refinement path change? Does refinement provide a larger reduction in error for queries with higher DEA uncertainty? Please also report failure-detection and selective-risk results using DEA itself.

2、Under comparable parameter budgets and the same refinement path and training settings, how does the proposed gate compare with a gate predicted directly from query features? As supplementary comparisons, how do an attention-entropy gate and a model trained with the Dirichlet mean instead of sampling perform? These comparisons would separate the effects of gating, stochastic training, and concentration-based uncertainty.

3、How are the Env-Split training and validation scenes distributed geographically? Please also describe how the keyword-derived condition labels were checked. Where map locations are shared between training and validation, does the performance advantage remain under location-matched evaluation? If too few scenes remain for this comparison, results reported separately for each map location would also be useful.

Reviewer_FQ5tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DUMAP introduced an evidence-allocation uncertainty modeling method in online vectorized HD Map reconstruction task. The authors chose Dirichlet distribution for this uncertainty modeling and Introduce a uncertainty fusion process called UGRF to fuse global uncertainty and local uncertainty. DUMAP also introduces a new data-split on nuScene to exclude the the adverse-condition data from training for construct the out-of-domain (OOD) experiment setting to test their proposed models. In experiments, DUMAP achives the SOTA in Env-Split and other data splits settings.

Critical strengths
  • The Env-split brings value since its more close to real world deployment.
  • The experiments and shows a clear advantages when applying DUMAP on OOD enviroments.
  • The uncertrainty is used for query aggregation and affecting the embedding which looks like a novel direct way to use it.
Critical weaknesses
  • There are lots works modeling uncertainty among queries (maybe not exactly on this specific task). Some comparison and discussion on why this new modeling is neccessary is missing.
  • Its hard to understand why Dirichlet Uncertainty and Coordinated Uncertainty are used at the same time in the same pipeline, while Coordinated Uncertainty shows a better sperman correlations (0.7+). If only use Coordinated Uncertainty would it remain the same performance? The ablation is missing for indepedent modules alone.
  • in Table 5, there could be a lot of exiting uncertainty signal could be selected for comparison and shaping the reasoning for choosing Dirichlet. Table 5 could be actually more important than Table 3 especially there is no ablation on performance of UGRF alone.
Questions for authors
  • Plese discuss the difference between "Map with Uncertainty Quantification for Robust BEV Vectorized Construction" (MapUQ) and DUMAP and why we need new uncertainty estimation methods in this tasks and showing
  • What would be the performance If only use Coordinated Uncertainty w/o EDA, the ablation seems incomplete. Would the coordinated be used as gate for query as well.
  • Its hard to understand if the improvement comes from the DEA uncertainty its self or just ramdom sampling regularity effects brought by the DEA, cause DEA shows no obvious sperman correlations to the actual error?
1 comment
AuthorsOpenReview ↗

Response to FQ5t: Q1, Q2

Thank you for asking how DUMap differs from MapUQ, how coordinate uncertainty relates to the fusion gate, and how much stochastic sampling contributes to performance.

W1. How DUMap Differs From MapUQ

MapUQ and DUMap

MapUQ and DUMap use uncertainty at different points in the mapping pipeline. MapUQ estimates aleatoric and epistemic uncertainty in BEV features with MC dropout and uses these estimates to weight semantic supervision. Its decoder adjusts the deformable-attention search region using geometric error during training and a point-dispersion proxy at inference. Its classification head also uses uncertainty for hard-negative mining.

DUMap models the normalized attention weights that each map query assigns to sampled BEV features. The Dirichlet mean gives the expected allocation, and the total concentration measures how tightly the allocation is distributed around that mean. UGRF uses this concentration-based signal to weight the local geometry and the globally refined geometry. This gives each query a measure of evidence-allocation reliability before geometric prediction.

What DEA adds

Why DEA is needed. Two queries can assign the same expected weights to sampled BEV features while having different allocation reliability. MapUQ’s uncertainty estimates address semantic prediction, search-region adaptation, and classification. DEA uses total concentration to characterize the reliability of the attention allocation itself before coordinate prediction, giving UGRF a signal for balancing local and global geometry.

We have cited MapUQ and discussed this distinction in Sec. 2 of the revised manuscript.

W2. Coordinate Uncertainty and the Fusion Gate

Coordinate-uncertainty baseline

The manuscript’s component ablation already reports this comparison: on Env-Split, the baseline with Laplace coordinate uncertainty and without DEA achieves 32.6 mAP. This uncertainty-aware MapTRv2 baseline retains the Laplace coordinate head and negative log-likelihood loss, with LF and UGRF disabled. Adding DEA alone raises mAP to 35.3 (+2.7 points).

Coordinate uncertainty as a gate input

DEA measures uncertainty in how each map query distributes attention over sampled BEV features. Its Dirichlet concentration is available before local–global fusion. UGRF uses it to give global refinement more weight when the query's local evidence allocation is less reliable. Coordinate uncertainty describes the precision of the predicted geometry. Each branch predicts a Laplace scale for each point and coordinate axis. The final coordinate scales are computed by combining the two branches' scales using the fusion weight, and are supervised by a Laplace negative log-likelihood on the coordinate residuals.

The fusion weight must be computed before the final coordinate uncertainty. Using the reported final coordinate uncertainty as the gate input would create a circular dependency: that uncertainty already depends on the very weight the gate is meant to compute. Its higher Spearman correlation with final geometric error (0.7181) shows that it ranks geometric errors well. UGRF maps each query’s DEA uncertainty to its fusion weight. The mapping is nondecreasing, so higher DEA uncertainty can increase the weight assigned to global refinement. For example, a prediction can remain uncertain when both branches struggle with the same occlusion, so high coordinate uncertainty alone does not tell us which branch to trust more.

The revised manuscript describes the two uncertainty signals and their roles in Sec. 3.4 and identifies the Laplace-only baseline in Sec. 5.1, Table 3.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revision Summary

We thank the reviewers for their constructive feedback. We revised the manuscript to clarify what DEA uncertainty measures, how it guides fusion, and how geographic overlap affects Env-Split evaluation.

  • DEA uncertainty under visual degradation. Section 5.2 and Appendix E.2 (Tables 19–20) report that controlled blur increases both DEA uncertainty and geometric error among instances detected before and after blurring. Across matched predictions, DEA uncertainty is positively associated with geometric error in all three adverse environments. The pooled Spearman correlation is 0.395 for DEA and 0.036 for the entropy of the mean attention weights.
  • Attention, fusion, and training controls. Section 5.1 and Appendix C (Tables 7–9) report the fixed-fusion attention comparisons, parameter counts, and alternative gates. Under fixed fusion, DEA achieves 38.10 mAP, compared with 35.63 for equal-parameter deterministic attention and 37.20 for Gaussian attention. The DEA gate achieves 39.25 mAP, and training with the Dirichlet mean achieves 39.23 mAP.
  • Branch errors and failure ranking. Appendix E.2 (Tables 21–22) shows that both branch errors increase across DEA uncertainty groups and that the global branch’s error reduction varies across groups. At 50% coverage, retaining the least uncertain predictions according to DEA reduces the observed failure rate from 46.5% to 19.9% and from 25.0% to 11.5% in the two prediction-score cohorts.
  • Env-Split labels and geographic overlap. Appendix D.2 (Tables 14–15) documents the review of the first image from each of the 846 scenes and the geographic screening protocol. DUMap leads UIGenMap by 2.8 mAP points on full Env-Split and by 0.54 points on the geographically disjoint 62-scene subset.
  • MapUQ and coordinate uncertainty. Section 2 clarifies how DUMap differs from MapUQ. Section 3.4 specifies that the final Laplace coordinate scale is computed after the fusion weight is determined, and Table 3 identifies the baseline that retains the Laplace coordinate head and its training loss.
AuthorsOpenReview ↗

Summary of Responses

Dear ICLR 2027 AC, SAC, and PC,

Thank you for considering our submission, DUMap (Paper ID 9895). We summarize the reviewers’ initial assessments and our responses below.

1. Core Contribution and Initial Reviewer Scores

DUMap models how each map query allocates attention over sampled BEV features. Dirichlet Evidence Allocation (DEA) uses the Dirichlet mean for inference-time aggregation, and Uncertainty-Guided Residual Fusion (UGRF) uses a concentration-derived signal to balance local predictions and global refinement. On nuScenes, DUMap achieves 39.2, 25.1, and 41.2 mAP on the Env-Split, City-Based, and Region-Based evaluations, respectively.

  • Reviewer UWbH · Rating 2 · Confidence 3. Asked about the validity of the DEA uncertainty signal, controlled comparisons with alternative attention and gate designs, and geographic overlap in Env-Split.
  • Reviewer fVjc · Rating 3 · Confidence 2. Asked about branch errors and failure ranking, alternative gates and mean training, and the validation of Env-Split labels and geography.
  • Reviewer FQ5t · Rating 2 · Confidence 3. Asked how DUMap differs from MapUQ, what coordinate uncertainty contributes, and whether stochastic sampling accounts for the performance gains.

2. Responses to Reviewer Concerns and Key Revisions

ConcernReviewer(s)Response and manuscript revision
Visual degradation and error correlationUWbH W1, FQ5t W3Section 5.2 and Appendix E.2, Tables 19–20 report the controlled blur and per-environment diagnostics. Among matched predictions, DEA–error Spearman correlations range from 0.126 to 0.359 across the three adverse environments. The pooled correlation is 0.395 for DEA and 0.036 for attention entropy.
Controlled attention comparisonsUWbH W2, FQ5t W3Section 5.1 and Appendix C, Tables 7–8 report 38.10 mAP for DEA, 35.63 for equal-parameter deterministic attention, and 37.20 for Gaussian attention under the same fixed fusion weight. DEA uses 7.35% fewer cross-attention parameters than Gaussian attention.
Alternative gates and stochastic samplingUWbH W2, fVjc W2, FQ5t W3Appendix C, Table 9 reports 39.25 mAP for the DEA gate, 37.72 for the direct query-feature gate, and 37.40 for the attention-entropy gate. Section 5.1 reports 39.23 mAP when training with the Dirichlet mean, within 0.02 points of sampling under the two-decimal comparison.
Branch errors and selective riskfVjc W1Appendix E.2, Tables 21–22 shows that the share of queries receiving greater global weight rises from 4.3% to 34.6%, while the global branch’s error reduction varies across uncertainty groups. DEA alone yields AUROC values of 0.842 and 0.757 in the two prediction-score cohorts.
Env-Split labels and geographic overlapUWbH W3, fVjc W3Appendix D.2, Tables 14–15 documents the label review and geographic screening. On Disjoint-62, DUMap reaches 36.83 mAP versus 36.29 for UIGenMap, with class AP computed on predictions pooled across three locations. The corresponding gain is 0.54 points on Disjoint-62 and 2.8 points on full Env-Split.
MapUQ and coordinate uncertaintyFQ5t W1–W2Section 2 distinguishes DUMap’s attention-allocation uncertainty from the uncertainty signals used by MapUQ. Section 3.4 clarifies that the final fused coordinate scale depends on the fusion weight. Table 3 reports 32.6 mAP for the baseline with Laplace coordinate uncertainty and 35.3 mAP after adding DEA.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.