- 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
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.
| Metric | TopoFG | TopoHG | Difference |
|---|---|---|---|
| Parameters | 70.85M | 73.20M | +2.35M (+3.32%) |
| Average latency (ms/frame) | 386.29 | 397.81 | +11.52 (+2.98%) |
| Peak allocated memory (MiB) | 2332.89 | 2341.86 | +8.97 (+0.38%) |
| FLOPs of profiler-supported operators (GFLOPs/frame) | 2542.20 | 2550.34 | +8.14 (+0.32%) |
The results show that our method introduces only a small number of additional parameters while achieving better results.
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:
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.