acceptodds
Under review as a conference paper at ICLR 2027

TopoHG: Reliable Hypergraph Reasoning for Lane Topology

Abstract

Inferring a structured road graph from surround-view images requires jointly de- tecting 3D lane centerlines and traffic elements, and predicting lane-to-lane and lane-to-traffic-element relations. Most query-based methods first detect individ- ual road elements and then predict topological relations for candidate element pairs. Under this paradigm, false-positive lane queries may still receive high confidence, allowing detection errors to propagate into topology prediction and induce spurious links. Moreover, pairwise prediction does not explicitly capture group-level topology patterns spanning multiple road elements, such as lane splits, merges, and multi-lane control associations. To address these limitations, we pro- pose TopoHG, a reliable-query-driven topology reasoning framework based on a structure-aware heterogeneous hypergraph. TopoHG combines contrastive query learning with online hard-negative mining to produce informative and discrimina- tive lane-query representations. The resulting reliability estimates softly weight lane queries during hypergraph construction, limiting the propagation of unreli- able candidates into topology reasoning. TopoHG then constructs structure-aware hypergraphs over lane and traffic-element queries, aligning with real-world split, merge, and control patterns. By integrating query semantics, lane geometry, and coarse pairwise predictions, cross-hyperedge interaction and gated bidirectional node–hyperedge message passing enable context sharing within and across struc- tural groups. The resulting group-level context is propagated back to individual queries to refine L2L and L2T relation predictions. On OpenLane-V2, TopoHG achieves OLS scores of 50.71 on Subset-A and 48.20 on Subset-B, with consistent gains across element detection and topology tasks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
3D lane topology reasoning, multi-view perception, hard-negative contrastive learning, heterogeneous hypergraph, autonomous driving

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

This paper proposes TopoHG for lane topology reasoning from surround-view images. The framework addresses unreliable lane queries and limited modeling of higher-order dependencies by combining reliability-aware query learning with a heterogeneous hypergraph. Contrastive supervision separates matched queries from high-confidence unmatched candidates, while reliability-weighted hyperedges encode lane splits, merges, and traffic-control associations to refine topology predictions. Experiments on OpenLane-V2 demonstrate improvements over the TopoFG baseline on both subsets.

Critical strengths

(+) The formulation is well motivated: unreliable detections can induce spurious relations, while split, merge, and control structures naturally involve multiple road elements. The two modules address complementary aspects of topology estimation.

(+) The independent ablations provide evidence of distinct contributions. RQL improves detection and topology metrics, whereas THG primarily improves topology: TOPll increases from 29.73 to 31.11 and TOPlt from 28.63 to 29.85 in Table 3.

Critical weaknesses

(-) The reliability head and detection classifier both use supervision derived from Hungarian matching, yet the paper does not establish what additional information the reliability estimate provides. A classification-confidence-only gate and measurements of false-positive relations would clarify its contribution.

(-) The current ablation studies are conducted only on TopoFG. Extending the component-wise ablations to other base models, such as TopoNet and TopoMLP, would provide stronger evidence of the method’s general applicability and clarify whether its gains depend on the specific underlying architecture.

(-)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

Could the authors provide the inference latency and memory measurements?

2 comments

1. Distinction Between the Reliability Estimation Head and the Detection Classifier

We thank the reviewer for the suggestion. The difference between them lies in their representation learning pathways and their usage in topological reasoning:

  • Different representation processing pathways. The detection classification head classifies individual point queries and then aggregates the classification logits. In contrast, the reliability branch first aggregates point-query features, then projects and normalizes them according to Eq. (1), and computes reliability scores through a separate prediction head.

  • The contrastive objective introduces explicit feature-relation constraints. Eq. (2) selects candidates from unmatched queries whose classification confidence exceeds a threshold as hard negatives, and performs contrastive learning against scene prototypes formed by matched queries.

  • The reliability head converts contrastive representations into scores for inference. According to Eq. (11) in the Appendix, the prediction head takes the normalized contrastive representations with detached gradients as input, so that the reliability loss does not alter the contrastive space through this readout pathway.

  • The fused score is used to control the participation weights in structural reasoning. Eq. (4) combines the reliability score with the single-class detection confidence, and Eq. (6) further uses the fused weight to determine hyperedge membership probabilities. Therefore, the purpose of this branch is to use the discriminative structure in the contrastive representations to regulate the strength of query participation in hypergraph aggregation, aiming to reduce the influence of low-reliability candidates on relation propagation.

Therefore, our method reuses the matching supervision and introduces additional contrastive constraints and representation readout to provide soft weights for hypergraph construction. Its design rationale comes from the differences in representation learning pathways.

2. Applicability of the Method to Other Backbones

To examine whether the proposed modules depend on TopoFG, we conducted standalone component ablations on the TopoLogic backbone. Using our locally reproduced TopoLogic as the baseline, we independently added RQL and THG. The results are as follows:

ConfigurationDET_l ↑DET_t ↑TOP_ll ↑TOP_lt ↑OLS ↑
TopoLogic Baseline29.26247.70123.94524.88743.946
+ RQL only29.84946.43924.99926.47944.434
+ THG only30.52648.36724.98425.70244.894

RQL improves the lane detection metric DET_l by 0.587 points and the two topology metrics by 1.054 and 1.592 points, respectively, resulting in an OLS improvement of 0.488 points. THG achieves improvements in all four component metrics: DET_l, DET_t, TOP_ll, and TOP_lt improve by 1.264, 0.666, 1.039, and 0.815 points, respectively, with an OLS improvement of 0.948 points. All the above gains are calculated relative to the original TopoLogic baseline. These results, together with the standalone ablations on TopoFG, provide complementary evidence that both modules support transferability to an additional backbone, indicating that their benefits are not limited to the specific implementation of TopoFG.

3. Evaluation on Downstream Tasks

We agree with the direction emphasized by AMap and Stability Under Scrutiny: the value of map predictions should be further examined in relation to downstream tasks and cross-frame stability.

  • The task scope of the benchmark used in this work needs to be clarified. OpenLane-V2 is primarily designed for perception and reasoning over lanes, traffic elements, and their topological relationships. Its original paper explicitly states that the dataset does not include planning tasks, because complete planning also requires information about dynamic participants such as vehicles and pedestrians. Therefore, conducting direct planning evaluation would require additional data alignment, a planner interface, and an evaluation pipeline.

  • Our outputs have downstream usage implications. L2L relationships provide a basis for lane reachability and path-branch selection, while L2T relationships provide a basis for associating lanes with traffic-control information. The modeling of split, merge, and control structures in TopoHG is precisely designed around these decision-relevant relationships. In this work, we first evaluate the prediction quality of these structural representations through a unified topology benchmark.

4. Inference Cost

We measured the inference cost of the complete model on a single RTX 3090 GPU. The complete TopoHG model was evaluated with a batch size of 1 over 200 frames. The average latency was 397.81 ms/frame, and the peak PyTorch allocated GPU memory was 2341.86 MiB. The latency includes model prediction and post-processing.

Inference Cost . We measured the complete TopoHG model on a single NVIDIA RTX 3090 GPU with a batch size of 1 over 200 frames. The model contains 73.20M parameters, with an average prediction and post-processing latency of 397.81 ms/frame and a peak PyTorch allocated GPU memory of 2341.86 MiB. The computational cost of operators supported by the PyTorch profiler for FLOP counting is 2550.34 GFLOPs/frame.

We further conducted a full-model comparison with TopoFG using the same input image resolution.Both models adopt a scaling setting of 0.75 for the front-view image and 0.5 for the other views, and use the same validation set, GPU, and batch size.

MetricTopoFGTopoHGDifference
Parameters70.85M73.20M+2.35M (+3.32%)
Average latency (ms/frame)386.29397.81+11.52 (+2.98%)
Peak allocated memory (MiB)2332.892341.86+8.97 (+0.38%)
FLOPs of profiler-supported operators (GFLOPs/frame)2542.202550.34+8.14 (+0.32%)

The results show that our method introduces only a small number of additional parameters while achieving better results.

Reviewer_XKSNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses lane topology reasoning from surround-view images, where a model jointly detects 3D lane centerlines and traffic elements and predicts lane-to-lane and lane-to-traffic-element relations. The authors identify two limitations of the detect-then-relate paradigm: unmatched query candidates that still receive high lane confidence can propagate into spurious topological links, and pairwise prediction decomposes structurally coupled group patterns such as splits, merges, and multi-lane control into independent binary decisions. The proposed method, TopoHG, combines Reliable Query Learning, which contrasts Hungarian-matched queries against a scene-conditioned prototype and high-confidence unmatched hard negatives in a separate projection space, using a gradient-scaling operator so the auxiliary signal does not disturb the shared features used for geometry regression, with Typed Heterogeneous Hypergraph reasoning, which organizes queries into split, merge, and control hyperedges weighted by the resulting reliability estimates and produces residual corrections to the base relation logits at the final decoder layer.

Critical strengths
  1. The problem is real and well-chosen. Detection errors propagating into topology prediction is a genuine failure mode in the detect-then-relate paradigm: a false-positive lane query that survives with high confidence becomes a node in the relation head and can only produce spurious edges. Acting on query reliability before relations are constructed, rather than filtering afterwards, is the right place to intervene.

  2. The limitations of prior work are analyzed precisely rather than generically. The paper does not simply assert that existing methods are insufficient. It states where CoPo’s contrastive objective operates (on relation embeddings after construction) and why that differs from suppressing unreliable queries beforehand, and it explains why pairwise decomposition cannot express the shared-anchor structure of splits, merges, and multi-lane control. This specificity makes the motivation checkable rather than rhetorical.

  3. The gradient-control design shows real care about a non-obvious conflict. Reliability learning and 3D geometry regression share the same point-query features, and an auxiliary contrastive gradient could plausibly degrade the fine-grained cues the regression needs. The paper anticipates this and uses an operator that leaves the forward pass unchanged while attenuating only the auxiliary gradient, and separately detaches the reliability gate so topology losses cannot reach back through that path.

Critical weaknesses
  1. Most of the reported gain comes from an image-resolution change, not from either proposed module. In Table 2, the front-view scaling adjustment alone accounts for roughly two thirds of the total improvement over the reproduced baseline, leaving the remainder split between RQL and THG. No result is reported for TopoFG under the same scaling, so the headline comparison cannot be attributed to the method.

  2. The hypergraph module, which is the paper’s main conceptual contribution, is not supported by its own ablation. In Table 3, THG alone yields a small OLS gain while both detection metrics drop. In Table 2, adding THG on top of RQL leaves TOP_ll slightly lower, even though split and merge hyperedges exist specifically to model L2L structure; the gain in that row comes mainly from an increase in traffic-element detection, which a topology-refinement module operating on relation logits has no stated mechanism to produce and which the paper does not explain. The three hyperedge types are also never ablated individually.

  3. The evaluation protocol does not separate tuning from reporting. All results, including every ablation, are on validation splits that were also used to select hyperparameters, while the test split each subset provides is never used. With a single run and no seed or variance reported, the margins are not separable from tuning and run-to-run noise. No hyperparameter is analyzed anywhere: the appendix supplies implementation details but no experiments, so the gradient-scaling coefficient on which the claim of protecting geometry-sensitive representations entirely rests is never varied, and the sparsification parameter determining hyperedge size is not assigned a value at all.

  4. Cost and limitations are not reported. The paper gives no parameter count, FLOPs, inference latency, or memory, which matters for a task intended to run onboard, and the hypergraph branch requires pairwise features over all directed lane pairs plus attention over several hundred candidate hyperedges in exchange for a small OLS gain. There is also no limitation or failure-case discussion, and several choices that appear to constrain the method, applying the hypergraph only at the final decoder layer and disabling its residual for the first third of training, are presented as implementation details without discussion of why they are necessary.

Questions for authors
  1. What does TopoFG achieve under the same front-view scaling? Without a matched-resolution baseline, the headline margin cannot be attributed to RQL or THG.

  2. Why does adding THG increase DET_t by 1.45 in Table 2, while TOP_ll slightly decreases? THG operates on relation logits with detached node features, so the pathway is unclear, yet this term drives the OLS gain in that row. Is the difference within run-to-run variation?

  3. How sensitive is the method to the gradient-scaling coefficient α? Only α = 0.1 is reported, and a comparison against different α would show whether the operator is necessary.

2 comments
  1. Response to Concern 1 We thank the reviewer for pointing out the input resolution when attributing the observed performance gains. We would like to clarify that the “+ Front-view scale” row in Table 2 corresponds to a reproduced TopoFG baseline with only the front-view resolution adjusted, without incorporating either RQL or THG. Therefore, this row already provides a baseline under the same input resolution, enabling a fair comparison with our full model.
  2. Response to Concern 2 We thank the reviewer' comment. Taken together, Tables 2 and 3 provide complementary experimental evidence supporting the topological benefits of THG. In Table 3, independently adding THG improves TOP_ll from 29.73 to 31.11 (+1.38) and TOP_lt from 28.63 to 29.85 (+1.22), resulting in an OLS improvement of +0.48. These results directly support the effectiveness of the hypergraph module in modeling both types of topological relationships. In Table 2, the −0.10 change in TOP_ll is measured relative to the intermediate configuration that already incorporates RQL. Compared with the resolution-matched TopoFG baseline, the full model still improves TOP_ll from 30.80 to 31.56, TOP_lt from 29.99 to 32.40, and OLS from 49.45 to 50.71. Thus, Table 2 demonstrates the overall improvement achieved by combining the two modules. Within this combined configuration, the additional contribution of THG is primarily reflected in L2T, while its ability to independently improve L2L is supported by the results in Table 3. The two modules may have overlapping effects, and improvements in individual metrics are not necessarily additive. Regarding DET_t, the feature separation blocks the direct gradient path from the hypergraph input to the node features. However, the base relation logits in Eq. (22) remain differentiable. The THG residual modifies the final relation predictions and their associated losses, which can consequently influence the optimization of the shared network through the base relation branch. Therefore, changes in detection performance after joint training have a plausible gradient-based explanation. We acknowledge that the +1.45 improvement in DET_t makes a substantial contribution to the OLS gain in Table 2. Nevertheless, evidence for the effectiveness of THG also comes from the independent improvements in both topological metrics in Table 3 and the further improvement in TOP_lt in Table 2. These results support the structural reasoning value of THG. The −0.10 change in TOP_ll in Table 2 alone is insufficient to attribute the overall performance gains to random variation across runs.
  3. Response to Concern 3 We thank the reviewer for this suggestion. Comparing different values of α would help further validate the role of gradient scaling. However, since each setting requires a complete retraining process, a systematic sensitivity analysis involves substantial computational cost and could not be completed within the current rebuttal period. Nevertheless, the direct effect of α is limited in scope. According to Eq. (3),Gα(q)G_{\alpha}(q) is an identity mapping in the forward pass and therefore does not alter the input features or inference computation. It only scales the auxiliary gradients from the contrastive loss flowing into the shared query representations, while leaving the direct gradients from the main task losses and the gradients of the projection head itself unchanged. The choice of α = 0.1 is intended to preserve the guidance provided by contrastive learning on the shared representations while reducing its potential interference with geometric regression.
  1. We thank the reviewer for their attention to the evaluation protocol. We would like to clarify that the validation-set evaluation setting adopted in this work is consistent with the public implementation of the baseline method TopoFG. The official repositories of TopoLogic, TopoFG, and other related methods explicitly report their main results on the OpenLane-V2 Subset-A validation set. Therefore, our comparisons and ablation studies are conducted on the same data split, following the public reproduction protocol of TopoFG, which ensures the comparability of the results. In addition, TopoMaskV3 also explicitly states that the official OpenLane-V2 test set is not publicly available.
  2. Inference Cost and Limitations. We measured the complete TopoHG model on a single NVIDIA RTX 3090 GPU with a batch size of 1 over 200 frames. The model contains 73.20M parameters, with an average prediction and post-processing latency of 397.81 ms/frame and a peak PyTorch allocated GPU memory of 2341.86 MiB. The computational cost of operators supported by the PyTorch profiler for FLOP counting is 2550.34 GFLOPs/frame.

We further conducted a full-model comparison with TopoFG using the same input image resolution.Both models adopt a scaling setting of 0.75 for the front-view image and 0.5 for the other views, and use the same validation set, GPU, and batch size.

MetricTopoFGTopoHGDifference
Parameters70.85M73.20M+2.35M (+3.32%)
Average latency (ms/frame)386.29397.81+11.52 (+2.98%)
Peak allocated memory (MiB)2332.892341.86+8.97 (+0.38%)
FLOPs of profiler-supported operators (GFLOPs/frame)2542.202550.34+8.14 (+0.32%)

The results show that our method introduces only a small number of additional parameters while achieving better results.

Reference:
TopoMaskV3: 3D Mask Head with Dense Offset and Height Predictions for Road Topology Understanding,2026.

Reviewer_YnN2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper propose TopoHG, a framework for lane topology prediction that aims to produce reliable lane-query learning with hypergraph-based reasoning. It uses contrastive learning to improve lane-query reliability and introduces typed hyperedges to model lane splits, merges, and traffic-control relationships. Experiments on OpenLane-V2 show improvements over existing methods.

Critical strengths

The motivation is interesting. it identifies an important challenges that complex road structures invovling multiple elements instead of pairwise relationships, which is the limitation of existing approaches.

Critical weaknesses
  1. I'm not fully convinced by the benefit of hypergraph reasoning. The paper argues that conventional pairwise relation modeling cannot well capture higher-order lane topology structures. However, the proposed hypergraph is still built from the original pairwise topology predictions, and the final outputs are still refined pairwise L2L and L2T scores. To me, this looks more like adding another message-passing module on top rather than introducing a fundamentally different way of modeling higher-order topology. I'm not sure what the hypergraph can capture that a regular GNN cannot. The whole design also feels a bit overcomplicated to me. I would like to see a comparison with a standard GNN refinement module, where all nodes can directly exchange information with each other without the extra selection using hypergraph construction. I think this would help to show whether the hypergraph construction is really needed.

  2. I'm also concerned about the extra computational cost. The hypergraph construction actually requires pairwise computations. In Eq. (5), the method compute split and merge membership scores for lane pairs and control membership scores for lane-to-traffic-element pairs, giving a complexity of O(NL2+NLNT)O(N_L^2 + N_LN_T). With 200 lane queries and 100 traffic-element queries, there are 100,000 candidate association scores per scene before top-K sparsification. However, the paper didn't report inference latency, FLOPs, or memory overhead, so it's hard to tell how expensive the proposed method actually is. I'm thinking whether the performance gains are really worth such a quadratic extra computation layer. It seems like the method is trading computational efficiency for better performance, but I'm not sure whether this trade-off is worthy. Especially for autonomous driving, the run-time efficiency matters a lot in real-world deployment.

Questions for authors

see the weakness above

3 comments
AuthorsOpenReview ↗

Response to concerns about hypergraph reasoning and computational cost

We thank the reviewer for raising these important concerns. We address the two issues separately below.

Concern 1: Distinction between hypergraph reasoning and conventional GNNs

1. Role of hypergraph reasoning

We agree that TopoHG uses the original pairwise predictions as one basis for constructing hyperedges, still outputs pairwise L2L and L2T scores. Our design aims to provide these predictions with reusable context associated with structural roles. Consider a split from lane A to lanes B and C. Eqs. (5)–(6) combine query features, geometry, original relation predictions, and reliability estimates to compute the membership weights of B and C in A’s split hyperedge. If both survive Top-K, Eqs. (15)–(16) (Appendix A.3–A.4) aggregate their features into a shared split representation[A→(B,C)]. Eqs. (19)–(20) (Appendix A.4) then refine A→B and A→C: both read A’s split representation while using their respective target-lane information. Thus, evidence from C can affect the A→B score. Merge and traffic-control hyperedges provide similar context. A fully connected GNN can also transmit such information. We do not claim that GNNs cannot capture higher-order relations; the distinction is that we explicitly construct and reuse split, merge, and control group representations.

2. Direct comparison with a fully connected GNN

To investigate whether explicit relational grouping provides advantages over general message passing, we conducted an additional comparison with a parameter-comparable Full-GNN baseline.

Under identical TopoFG, resolution adjustment, and RQL configurations, Full-GNN enables direct interactions among all lane and traffic-element nodes without constructing hyperedges or applying Top-K selection. Both models were trained for 24 epochs.

Table X. Full-GNN vs. TopoHG.

MethodOLSTOP_llTOP_lt
Full-GNN49.8531.4430.76
TopoHG50.7131.5632.40
Δ+0.86+0.12+1.64

TopoHG improves OLS by 0.86 and TOP_lt by 1.64 compared with Full-GNN, while improving TOP_ll by 0.12.

We also provide qualitative visualizations in the Supplementary Material (Sec. A.6, Fig. 4, p. 16). The figure compares TopoFG, Full-GNN, TopoHG, and ground truth under the same evaluation setting. In the upper-row scene, Full-GNN fails to recover the right-side turning branch and its associated connection, whereas TopoHG recovers a branch structure closer to GT. In the lower-row scene, Full-GNN misses the upper-right turning connection, while TopoHG preserves it and reduces some redundant branches. These matched examples provide complementary qualitative evidence that explicitly organizing relational groups can yield more complete local topology predictions than generic all-node message passing under the same configuration. We will revise overly strong statements, such as "pairwise modeling cannot capture higher-order structures," to clarify this distinction.


Concern 2: Computational complexity and inference overhead

1. Complexity of association computation

We agree that Eq. (5) requires computing 2NL2+NLNT2N_L^2+N_LN_T association scores before Top-K selection. With NL=200N_L=200 lane queries and NT=100N_T=100 traffic-element queries, this corresponds to 100,000 association scores. Top-K selection does not reduce this initial computation.

In our implementation, query features are first projected and association matrices are computed using batched operations on 128-dimensional features. This process does not require additional executions of the image backbone or BEV decoder for individual candidate pairs. Furthermore, hypergraph reasoning is performed only at the final decoder layer. Therefore, the pairwise computation introduces only a small incremental cost to the complete inference pipeline.

2. Measured inference overhead

To quantify the practical overhead, we conducted measurements on an NVIDIA RTX 3090 using the same final model configuration and checkpoint, with only the hypergraph branch enabled or disabled.

Each setting was measured twice with batch size 1 and 200 frames per run.

Table Y. THG overhead.

MetricOffOnΔ
Latency (ms/frame)394.05398.72+4.67 (+1.19%)
Params71.02M73.20M+2.19M (+3.08%)
Peak memory (MiB)2333.522341.86+8.34 (+0.36%)
Profiler FLOPs (GFLOPs)2542.592550.34+7.74 (+0.30%)

3. Performance–efficiency trade-off

As reported in Table 2, adding THG improves OLS from 50.13 to 50.71 (+0.58).

Together with the measured latency and memory overhead, these results indicate that THG achieves a favorable performance–efficiency trade-off under our current experimental setting.

Thanks for providing the results. Considering the moderate difference between the fully connected GNN and the hypergraph design (one of your core novelties), and the heavy inference overhead (with quadratic complexity) trading off against the OLS increment of 0.58, I think it has limited contribution to the field. I will maintain my score.

We thank the reviewer for this comment. The computational complexity introduced by the hypergraph component in TopoHG is (2N_L^2+N_LN_T). However, the original fully connected GNN also has quadratic complexity, i.e., ((N_L+N_T)^2). Therefore, the additional computational complexity of TopoHG over the GNN is (N_L^2-N_T^2-N_LN_T). In practice, under the current query configuration, pairwise associations are computed in batches using projected 128-dimensional features only at the last decoder layer. This avoids repeatedly executing the image backbone or BEV decoder for each candidate pair. Consequently, the measured latency overhead of the hypergraph branch is relatively small compared with the overall inference pipeline. Our results show that enabling the hypergraph branch increases inference latency by only approximately 1% (see Table Y). Meanwhile, TopoHG achieves an overall OLS improvement of 0.86 (approximately 1.72%), accompanied by a 1.64-point improvement in TOP_lt (see Table X). These improvements are consistent with the qualitative results shown in Fig. 4 of the Appendix, which demonstrate the benefits of TopoHG in intersection scenarios. The improvements are particularly evident in the successful prediction of additional lane relationships at intersections. Since intersections typically involve more traffic elements and more complex topological relationships, improving lane detection and topology reasoning in such scenarios is particularly important.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.