acceptodds
Withdrawn from ICLR 2027

Refine Connections, Close the Gap: A Reliable Enhancement Framework for Driving Scene Topology

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In autonomous driving, understanding scene topology—the connectivity between lanes and traffic elements—is critical for safe path planning and motion control. While current methods excel at detecting individual map elements, their connectivity reasoning often falls short of its theoretical potential, leaving a significant performance gap relative to the theoretical upper-bound (Oracle) achievable given the underlying detections. Furthermore, the decision-ready topology graphs passed to downstream tasks often remain unreliable. Current approaches typically derive connectivity by thresholding continuous topology scores; however, these scores often fail to reflect the true logical likelihood of connectivity, resulting in false positives or missing connections. Existing benchmarks further overlook this issue by primarily evaluating continuous metrics (e.g., score), rather than assessing the discrete connectivity required for decision-making. To bridge these gaps, we propose TopoEnhance, a novel topology enhancement framework designed to unlock the latent potential of existing methods and improve the reliability of decision-ready topology. We formulate topology enhancement as a denoising-based reconstruction process, where the model learns to recover structural consistency from stochastically corrupted ground-truth graphs. This formulation enables the model to resolve logical inconsistencies and rectify unreliable connections, producing robust discrete topology graphs that closely approach theoretical maximum performance. Extensive experiments across different baselines show that TopoEnhance consistently improves both continuous topology metrics ( and ), and discrete connectivity measured by our adapted Topology Jaccard Similarity () metric. As a flexible, source-agnostic framework, TopoEnhance delivers substantial gains across diverse state-of-the-art baselines without requiring retraining.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Driving Scene Topology

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_sczwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TopoEnhance, a framework for improving topology connections in driving scenes. The authors point out that existing methods still have a considerable performance gap compared with their theoretical upper bounds, and that the quality of the final discrete topology graphs is not sufficiently evaluated. To address these issues, the authors introduce a topology noise simulation strategy and train a heterogeneous GNN to refine connectivity predictions. The refined scores are then combined with the original predictions without retraining the baseline models. The paper also introduces Topology Jaccard Similarity (TJS) to evaluate discrete connections. Experiments on OpenLane-V2 show consistent improvements across several existing methods.

Critical strengths
  1. The paper addresses a meaningful problem in driving scene topology reasoning. Reliable topology connections are important for downstream tasks such as path planning. The authors highlight the need to evaluate the quality of the final discrete topology graphs, which is not fully reflected by commonly reported topology metrics. Introducing TJS as a complementary evaluation metric is a useful contribution.
  2. The experimental results are promising. TopoEnhance improves both TOP and TJS across several baselines, including TopoNet, TopoMLP, TopoLogic, and SMART. The improvements in TJS are particularly substantial. These results demonstrate the potential of improving topology predictions without retraining the original models.
  3. The paper provides useful additional experiments beyond the standard benchmark results. The evaluations on geographical generalization, temporal consistency, safety-related topology errors, and route recoverability help demonstrate the potential practical value of the proposed method.
Critical weaknesses
  1. The methodological novelty is not sufficiently clear. The proposed framework mainly combines noise augmentation, a heterogeneous GNN, and score fusion, which are relatively established techniques. Previous works such as TopoLogic have also explored improving topology predictions through geometric information and post-processing. Although the proposed framework achieves promising results, its main methodological contribution over existing approaches needs to be better justified. Comparisons with simpler refinement methods would help clarify this point.
  2. The large improvements in TJS require further analysis. The paper uses a fixed threshold of 0.5 to convert topology scores into discrete connections. However, different baseline models may produce scores with different distributions. As a result, the relatively low TJS scores of the original models may partly result from the threshold choice rather than poor topology reasoning alone. It is therefore unclear how much of the reported improvement comes from better structural predictions and how much comes from changes in score calibration. Comparisons with baselines using individually tuned thresholds or simple calibration methods would help address this concern.
  3. The contribution of the proposed noise simulation strategy is not sufficiently isolated. It remains unclear whether the performance improvements mainly come from the noise simulation and denoising formulation, or from the additional GNN trained to predict topology connections. A comparison with a standard supervised GNN trained without the proposed noise simulation would help establish the necessity of the denoising design.
  4. The distinction between TOP and TJS could be explained more clearly. The paper describes TOP as a metric for continuous topology scores and TJS as a metric for discrete connections. However, the official OpenLane-V2 implementation also applies confidence thresholding when computing topology AP. A more precise discussion of the differences between these metrics, particularly regarding thresholding and the evaluation of discrete connections, would strengthen the motivation for introducing TJS.
Questions for authors
  1. How sensitive are the reported TJS results to the confidence threshold? Could the authors report results under different thresholds and compare against baselines with thresholds tuned on a separate validation split?
  2. How does TopoEnhance compare with simpler approaches, such as score calibration or geometry-based post-processing? Such comparisons would help clarify the benefits of the proposed framework over existing alternatives.
  3. Could the authors provide an ablation study using a standard supervised GNN without the proposed noise simulation strategy? This would help isolate the contribution of the denoising formulation.
  4. Could the authors clarify the differences between TOP and TJS, particularly regarding confidence thresholding and the evaluation of discrete topology connections?
Reviewer_vjA6OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

TopoEnhance refines lane-lane and lane-traffic connectivity while keeping upstream detections fixed. It trains a heterogeneous graph attention network on synthetically corrupted ground-truth graphs using relation-normalized binary cross-entropy. The refined scores are fused with the original topology predictions. Experiments on OpenLane-V2 report topology ranking (TOP), thresholded graph quality (TJS), and supplementary analyses of geographic generalization, temporal consistency, route recoverability, and efficiency.

Critical strengths

The module improves thresholded topology quality across several prediction sources without retraining their detectors. The geographic-split and route-recoverability experiments extend the evaluation beyond aggregate topology scores.

Critical weaknesses
  1. The synthetic-to-real distributional claim is unsupported. Section 3.2 (pp. 4-5) claims that q(G_noise | G*) closely matches real prediction distributions without defining closeness or measuring agreement. It also compares a conditional corruption distribution with a marginal prediction distribution without specifying the required conditioning or marginalization. Algorithm 1 perturbs ground-truth instances and adds false positives, but provides no explicit mechanism for simulating missed detections. Matching expected node counts cannot establish agreement in recall, correlated localization errors, or connectivity structure. Table 9 tests artificial score flips and provides no direct evidence for the claimed distributional match. The paper therefore does not substantiate the explanation on which its generalization argument relies.

  2. The specified refinement decoder cannot represent directed connectivity. Equation (5), p. 6, computes sigmoid(cos(h_u,h_v)) from the same node embeddings for both endpoint roles. This score satisfies s_uv = s_vu, whereas lane successor relations are directed. Directional message passing cannot remove the symmetry imposed by the final scoring function. The specified refinement branch therefore cannot distinguish u->v from v->u. The authors must provide the actual directional scoring mechanism and reconcile it with Equation (5). This discrepancy directly affects the validity and reproducibility of the method description.

  3. Missing calibration controls leave the main gains unexplained. The paper evaluates TJS at a fixed threshold of 0.5 without comparing against validation-tuned thresholds or simple calibration of the original scores. It therefore does not establish how much improvement requires structural denoising. Section 4.1 also conflates ranking improvement with probability calibration. A strictly increasing score transformation preserves ranking but can change TOP, since its official implementation applies a confidence threshold of 0.5 before computing AP. Higher TOP alone does not establish better probability calibration. Appendix C further gives lane-lane fusion weights ranging from 0.6 to 0.9 without specifying the settings and selection procedure for each prediction source. These omissions prevent a controlled and reproducible assessment of improvements across prediction sources.

  4. Evidence on stronger recent methods is inadequate. The evaluation omits RATopo, TopoFG, and CoPo, leaving the broad applicability claim insufficiently tested. For TopoPoint, Table 4 shows TOP_ll gains of only 0.11 on Subset_A and 0.27 on Subset_B. These results provide little evidence of substantial lane-lane ranking improvement on a stronger prediction source. The Subset_B reproduction also falls below published TopoPoint performance, so it does not test refinement of the predictions of the published system. The evaluation needs to establish whether the proposed contribution remains significant as upstream topology reasoning improves.

  5. The ablations do not establish the necessity of the proposed design. Appendix D.4 tests visual backbones, loss variants, and geometric jitter, mainly using TopoNet on Subset_A. Removing jitter changes TOP_ll by only 0.1 and TOP_lt by 0.4 (Table 8). The experiments do not isolate false-positive synthesis or the adaptive noise curriculum, despite their central role in the proposed corruption process. There is also no matched-capacity comparison against a simpler pairwise predictor to establish the value of heterogeneous message passing.

Minor Issues

  1. Numerical and efficiency reporting is inconsistent. Table 6 mixes decimal precision within individual columns, including 20.3, 21.8, and 21.84. Its accompanying text states that the performance gap between DINOv2-ViT-L and ResNet-50 is 0.2, without naming the metric. Their displayed TOP_lt values differ by 0.3 (25.8 versus 25.5), and their TOP_ll values differ by 0.04 (21.84 versus 21.8). Neither difference matches the text. Table 14's latency components sum to 6.06 ms, but the reported total is 6.43 ms. Score prediction takes 0.64 ms in the table and 1.02 ms in the text. Appendix C reports 0.1 ms/sample, while Appendix E.5 reports 6.43 ms for a dense graph. The graph sizes, timing scope, unlisted overhead, and inclusion of visual feature extraction must be specified to make the efficiency claims verifiable.
Questions for authors

It is suggested that the authors provide further clarification and supporting evidence to address the weaknesses discussed above.

Reviewer_ZhjEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TopoEnhance, a post-processing framework for driving-scene topology. A heterogeneous GNN learns connectivity from synthetically corrupted ground-truth detections and fuses its predictions with existing topology scores. The paper emphasizes the distinction between ranking quality and thresholded graph quality, introduces a topology Jaccard metric, and reports improvements across multiple OpenLane-V2 baselines. Additional analyses examine geographic generalization, temporal consistency, and route recoverability.

Critical strengths

(+) Improving connectivity while holding detections fixed is a useful and clearly motivated problem. The detection-conditioned oracle helps distinguish errors that a topology enhancer can correct from limitations inherited from the upstream detector.

(+) The gains span several upstream methods and both dataset subsets. For example, SMART improves from 37.0 to 40.1 TOPll and from 38.1 to 87.7 TJSll on Subset-A.

Critical weaknesses

(-) TopoEnhance refines connectivity while keeping upstream detections fixed. Although Appendix D.3 reports improvements in the overall OLS, approaching the detection-conditioned oracle does not necessarily imply accurate recovery of the complete scene topology, particularly when critical lanes or traffic elements are missed. Could the authors quantify how the refinement gains vary with detection quality, especially under missed detections and geometric localization errors, and clarify the limitations of connectivity-only refinement?

(-) Although Appendix E.5 reports latency and memory overhead, the measurement scope needs clarification. Do the reported 6.43 ms latency and 0.3 MB additional VRAM include the DINOv2-ViT-L feature extractor, or only the refinement module operating on precomputed features? Please report the complete incremental inference cost, including visual feature extraction, model parameters, and peak GPU memory under a clearly specified setup. Please also reconcile the 0.1 ms figure in Appendix C with Table 14 and resolve the discrepancies between the component-wise and total latency values.

(-) Recent work on online mapping has begun to examine how map predictions affect downstream tasks. Evaluating the proposed method on downstream tasks such as planning would help establish whether improved topology reasoning translates into practical benefits and would strengthen the paper’s empirical contribution.

[1] AMap: Distilling Future Priors for Ahead-Aware Online HD Map Construction, CVPR 2026

[2] Stability Under Scrutiny: Benchmarking Representation Paradigms for Online HD Mapping, ICLR 2026

Questions for authors

Q1: Appendix C reports 0.1 ms per sample, whereas Table 14 measures 6.43 ms for a dense stress-test graph. Please specify the workloads and batching behind each figure, and whether timing includes DINOv2 inference or only projection of precomputed features.

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

Related papers

Discussion (0)

Sign in to comment.