TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes

As an emerging task that integrates perception and reasoning, topology reasoning in autonomous driving scenes has recently garnered widespread attention. However, existing work often emphasizes "perception over reasoning": they typically boost reasoning performance by enhancing the perception of lanes and directly adopt MLP to learn lane topology from lane query. This paradigm overlooks the geometric features intrinsic to the lanes themselves and are prone to being influenced by inherent endpoint shifts in lane detection. To tackle this issue, we propose an interpretable method for lane topology reasoning based on lane geometric distance and lane query similarity, named TopoLogic. This method mitigates the impact of endpoint shifts in geometric space, and introduces explicit similarity calculation in semantic space as a complement. By integrating results from both spaces, our methods provides more comprehensive information for lane topology. Ultimately, our approach significantly outperforms the existing state-of-the-art methods on the mainstream benchmark OpenLane-V2 (23.9 v.s. 10.9 in TOP$_{ll}$ and 44.1 v.s. 39.8 in OLS on subset_A. Additionally, our proposed geometric distance topology reasoning method can be incorporated into well-trained models without re-training, significantly boost the performance of lane topology reasoning. The code is released at https://github.com/Franpin/TopoLogic.

Paper

References (41)

Scroll for more · 29 remaining

Similar papers

Peer review

Reviewer E2Wq4/10 · confidence 4/52024-07-09

Summary

The paper investigates topological reasoning for lane graph prediction for autonomous driving. The paper focuses on improving the prediction of the topological structure of the lane graph. Therefore, they propose two new mechanisms, the first is a geometric approach that estimates connectivity based on the distance between the end point and starting point of center lines. Secondly, they propose a semantic approach which uses the similarities of lane queries in a high dimensional space. These two approaches stand in contrast to existing works which mainly use MLPs to predict the connectivity of the lanes. They show how these approaches can be interleaved into a lane graph network and how it improves results, especially the topological reasoning results. Even adding only the geometric approach to existing methods improves the topological reasoning results.

Strengths

The paper is well written and easy to follow. The paper and results highlight the importance of geometric cues when predicting conductivity. Or in other words, the fact that the geometric approach works so well shows that there is a lot of potential in current methods. Furthermore, the geometric approach can be integrated into existing networks or combined with other approaches and allows for competitive lane topology prediction. Both the geometric and semantic prediction branches are clear and make sense and are able to improve results for topological metrics.

Weaknesses

- It is not clear to me why the paper does not compare with TopoMLP, the numbers for Table 2 v1.0.0 can just be taken out of the paper. - When comparing with TOP_lt TopoMLP outperforms TopoLogic and overall performance is very similar, in this case both methods use the same backbone. - The paper highlights that geometric cues in predicting the topology of the lane graph are underutilized which is interesting, but I am not fully sure if this is a Neurips paper, since technical contributions are marginal. - See questions

Questions

- Given that f_ours is just a gaussian function with an adapted decision boundary. I think the fair comparison would be with a hand tuned decision boundary, for example by looking at end point errors of a trained model. - f_tanh in equation 6 seems to map to [0, 2] and not to [0, 1] - Equation 10 is not a [0, 1] probability anymore, to remain this property you would need to take a convex combination of the two values. Do you perform some sort of normalization? - If you say a variable is 1x1 do you mean it is a real number? If so just write \in R - The batch size of 2 on 8 GPUs seems strange, is it a batch size of 2 per GPU, so an actual batch size of 16? - How do you handle traffic elements? - Typos: - L90: . some methods - not capitalized after full stop - L204: Both models - three models are listed before - L236: ),especially - L259: Table 4 Indicate show

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Has been discussed

Reviewer SnEK4/10 · confidence 4/52024-07-09

Summary

Reasoning about lane topology is becoming more and more important for the autonomous driving community. Current methods mostly focus on improving the perception performance and ignore the reasoning part of the task. The authors raise the importance of the relationship between lanes, namely the lane distance and lane similarity. The lane distance is determined by the geometry distance between the starting and ending points of the lanes. Then, the connectivity is decided based on the result of a learnable threshold function. The lane similarity is calculated based on the lane queries, and a large similarity indicates a large possibility of connectivity. Both lane distance and lane similarity modules serve as plug-in modules to existing networks. Experimentally, both modules have significantly improved the existing methods.

Strengths

- The authors observed that current methods on lane topology reasoning mostly focus on improving perception accuracy but ignore topology reasoning. The proposed modules serve as a good posting-processing technique for topology reasoning. - The proposed modules are effective and can be placed in existing networks.

Weaknesses

- In the experiment session, the authors provide scores on lane segment prediction. However, the lane segment is different from the lane centerline. Only the lane centerline is discussed in the methodology session. - It is not stated which split of the dataset is used for experiments. - The mixture of using different versions of the metric is confusing (vx.0 and vx.1). As stated in the repository of the dataset, this difference leads to different TOP scores. For instance, in Table, metrics in v2.0 and v2.1 are both listed, which is unreasonable. - The two proposed modules focus on topology reasoning. However, experimentally, the detection scores of the proposed pipeline also improved compared to other methods. This improvement is not discussed.

Questions

- The two proposed modules are plug-in modules on current networks, and the proposed pipeline is trained in an end-to-end manner. However, as these two modules focus on topology reasoning, and the final topology prediction is decided by these two modules, it is curious to know how these two modules benefit existing networks. Namely, only parameters in the proposed modules are trained, and the remaining weights of the existing networks are frozen. - For the second proposed module, the lane similarity module. Logically, the lane queries usually encode positional information of lanes. It is curious to know why improving the similarity between two connected lanes would improve topology reasoning performance. That is, the only similar thing between two connected lanes is the ending point of one lane and the starting point of another. Two parallel lanes should have a larger similarity than two connected lanes.

Rating

4

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

N/A

Authorsrebuttal2024-08-09

**Thank you very much for your quick response.** **Q7:** For Q2, are the scores reported on the validation set or test set? **A7:** We're sorry we didn't mention that detail, and we will add more clarification in the final version. In fact, all the results are reported on the **validation set**, which is **completely consistent** with the protocol of the **all reference paper**, e.g., TopoNet, SMERF, LanSegNet, etc. As for test set, it is only used for the closed-evaluation of the OpenLane-V2's challenge(https://opendrivelab.com/challenge2024/#mapless_driving), which was not available at the time of our research. ---- **Q8:** For Q5, the DET$_t$ of the frozen TopoNet drops compared to the baseline version, are there modules which could affect the traffic element detection branch? **A8:** We're sorry to confuse you. Either similarity module or geometric distance module just involves the **lane**, not **traffic elements**, so DET$_t$ really shouldn't drop. However, in the freezing experiment we loaded the weights of the TopoNet that we reproduced, whose DET$_t$ in itself is lower (**46.9** v.s. **48.6**) than reported in the TopoNet paper. It seems to be **a bug in baseline's official code**. The reproduction about DET$_t$ was always **unstable and lower** than the value reported in the original paper, and this phenomenon can also be observed in the results of other paper (e.g., **44.5** DET$_t$ of TopoNet in Table 1 in paper SMERF$^{[1]}$) [1] Luo K Z, Weng X, Wang Y, et al. Augmenting lane perception and topology understanding with standard definition navigation maps[J]. arXiv preprint arXiv:2311.04079, 2023. **If you have any further questions or suggestions about our article, we'd love to discuss our content with you.**

Reviewer FMUS5/10 · confidence 3/52024-07-12

Summary

The authors claim previous topology reasoning methods typically boost reasoning performance by enhancing the perception of lanes and directly adopt MLP to learn lane topology from lane query. In this work, the authors propose to make full use of lane geometric distance and lane query similarity for topology reasoning. The proposed method, named TopoLogic, mitigates the impact of endpoint shifts in geometric space, and introduces explicit similarity calculation in semantic space as a complement. The method achieves SOTA results in OpenLane-V2 dataset.

Strengths

1. The idea of paying more attention on geometry and similarity to help topology reasoning makes sense. It makes the reasoning process more interpretable and robust. 2. The performance gain brought by the proposed method is significant. The authors provide adequate ablation experiments to validate the design. 3. Many qualitative results are presented to show the effectiveness of the proposed method. The conection between lane endpoints are obviously improved.

Weaknesses

1. Like an incremental work based on TopoNet. The contribution is limited. The contributed part is only a similarity module for mapping lane line endpoints. But the design is simple and straightforward, not surprising enough. 2. The proposed GeoDist module is a post-processing algorithm. It's like an engineering improvement rather than academic contribution.

Questions

When compared with previous methods (like TopoNet, SMERF), what about the computation overhead and the difference in network architecture?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Refer to the Weaknesses part.

Reviewer 8GTK5/10 · confidence 4/52024-07-16

Summary

This paper proposes an interpretable method for lane topology reasoning based on lane geometric distance and lane query similarity. The authors reveal that the lane topology is easily disturbed by the endpoint shifts. Based on this, the proposed post-processing module improves the robustness and performance of lane topology reasoning, which can be plugged into other methods without re-training. And it also alleviates the influence of inaccuracies in lane detection on topology reasoning. Extensive experiments on OpenLane-V2 demonstrate its state-of-the-art performance.

Strengths

- The authors reveal the phenomenon that topology reasoning is highly susceptible to lane line endpoint shifts, providing a solid foundation and motivation for subsequent research. - The proposed post-processing module improves the robustness and performance of lane topology reasoning, which can be plugged into other methods without retraining. - The entire paper is easy to understand. The lane graph in Figure 5 shows very clearly the topological relationships between lanes and the differences from previous methods, which should be followed by other subsequent work.

Weaknesses

- The proposed method is relatively simple and involves only post-processing modules, which at the same time leads to limited contributions and incremental improvements. - Considering that TopoNet is just a simple baseline model, the boost in topological reasoning is not surprising. Moreover, as can be seen in Table 1, the improvement of $TOP_{lt}$ is very small and $DET_{t}$ even decreases, which needs further analysis and discussion.

Questions

- As can be seen in Figure 3, compared to other mapping functions, the final result of the learnable Gaussian mapping function has a much higher threshold. Therefore, compared to the present comparison experiments ($f_{gau}, f_{sig}, f_{tan}$), it may be better to manually set the hyperparameters for a series of gradients to observe its effect on the final performance. And I'm wondering why $f_{gau}, f_{sig}, f_{tan}$ are very close but differ dramatically on $TOP_{ll}$ in Table 3. - According to L146-149, the motivation of Lane Similarity Topology is to alleviate the influence of inaccuracies in lane detection on topology reasoning. However, the proposed method as a post-processing step does not significantly change the lane line detection results. In addition, the comparison results of the Lane Similarity Topology module should be added to Figure 4 to increase its persuasiveness. There are many typos in the paper, which should be further checked and polished. For example: L6: "are prone to" -> "is prone to" L12: "our methods provides" -> "our methods provide" L17: "boost" -> "boosting" Figure 1 caption: "reasoing"

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

This paper is limited in the post-processing for topology reasoning.

Reviewer SnEK2024-08-08

Thank you for your feedback. I still got some question. For Q2, are the scores reported on the validation set or test set? For Q5, the DET_t of the frozen TopoNet drops compared to the baseline version, are there modules which could affect the traffic element detection branch?

Reviewer FMUS2024-08-12

Further questions

Thanks for the authors' feekback. My further questions are as follows: Why does TopoNet-based TopoLogic and SMERF-based TopoLogic run faster and have less params and flops compared with original TopoNet and SMERF? Is the post-processing step (GeoDist module) included?

Authorsrebuttal2024-08-12

**Thank you for your insightful question.** **Q4**: Why does TopoNet-based TopoLogic and SMERF-based TopoLogic run faster and have less params and flops compared with original TopoNet and SMERF? **A4**: TopoNet and SMERF both compute lane topology directly using MLP. In contrast, TopoLogic calculates lane topology through a geometric distance module and a similarity module. This distinction highlights the structural differences between TopoLogic and both TopoNet and SMERF. In the TopoNet and SMERF, two MLPs are used to generate $Q_{emb} \in \mathbb{R}^{N \times C}$, where $N$ is the number of query. Subsequently, $Q_{emb}$ are repeated to $Q_{emb}^{'} \in \mathbb{R}^{N \times N \times C}$. **They are then concatenated and processed through $\operatorname{MLP_3}$ to compute the lane topology**, as described by the following Equations: $$ Q_{emb_1}, Q_{emb_2}=\operatorname{MLP_1}(Q_l^i), \operatorname{MLP_2}(Q_l^i) \in \mathbb{R}^{N \times C} $$ $$ Q_{emb_1}^{'},Q_{emb_2}^{'}=\operatorname{Repeat}(Q_{emb_1}),\operatorname{Repeat}(Q_{emb_2}) \in \mathbb{R}^{N \times N \times C} $$ $$ \operatorname{G_{sim}}=\operatorname{MLP_3}(\operatorname{Concat}(Q_{emb_1}^{'}, Q_{emb_2}^{'})) \in \mathbb{R}^{N \times N} $$ For TopoLogic, we calculates lane topology through a geometric distance module and a similarity module. In similarity module, We directly compute lane topology via matrix multiplation between $Q_{emb_1}$ and the transposition of $Q_{emb_2}$, followed by applying a sigmoid activation function, as detailed in our paper: $$ Q_{emb_1}, Q_{emb_2}=\operatorname{MLP_1}(Q_l^i), \operatorname{MLP_2}(Q_l^i) \in \mathbb{R}^{N \times C} $$ $$ \operatorname{S}=\operatorname{matmul}(Q_{emb_1},\operatorname{transpose}(Q_{emb_2})) \in \mathbb{R}^{N \times N} $$ $$ \operatorname{G_{sim}}=\operatorname{sigmoid}(\operatorname{S}) \in \mathbb{R}^{N \times N}\\ $$ In geometric distance module, as illustrated in Equation (4) of our paper, we utilize only two parameters, $\alpha$ and $\lambda$. Additionally, the merge process involves two parameters, $\lambda_1$ and $\lambda_2$, as shown in Equation (10) of our paper. All these parameters are scalar values, i.e., $\alpha,\lambda, \lambda_1, \lambda_2\in\mathbb{R}$. To summarize, while TopoNet and SMERF models involve **three MLPs** with associated parameters, TopoLogic utilize only **two MLPs** in the similarity module, two parameters $\alpha$, $\lambda$ in geometric distance m odule, and $\lambda_1$ and $\lambda_2$ in merge process. **In TopoLogic, the scalar parameters in the geometric distance module are significantly fewer in number compared to the MLP$_3$ parameters in TopoNet and SMERF.** Additionally, the MLP$_3$ results in a higher computational complexity. Therefore, TopoLogic has a smaller number of parameters (#PARAMS) and lower computational complexity (FLOPS), leading to faster speeds (FPS) compared to TopoNet and SMERF. **We hope this explanation addresses your concern, thank you again for your valuable feedback.**

Reviewer 8GTK2024-08-12

Thanks for the authors' positive feedback. It seems that the results of $TOP_{ll}$ are very sensitive to the hyperparameters. And as shown in Figure 3 of the global rebuttal, the introduction of the similarity module removes many predictions of topology. But both results of (b) and (c) are far away from GT. I think that the existing metrics may not reflect the results well, and the filtering could lead to large variations. In addition, the authors are not able to clearly explain the results of the $DET$. The above makes the quantitative experiment less convincing. Overall, I still have concerns about incremental improvements and contributions, which are also mentioned by other reviewers.

Authorsrebuttal2024-08-12

**We are very grateful that you have carefully reviewed our rebuttal response.** Although there are deficiencies in the details, we still believe that this work has sufficient contributions to the field: **we have significantly improved the core metrics (TOP$_{ll}$ & OLS) of the emerging task of topology reasoning on autonomous driving with a easily understandable method.** Regarding your further questions, we will answer them one by one below. We hope you can reconsider the rating of our article. ---- **Q6.1:** It seems that the results of TOP$_{ll}$ are very sensitive to the hyperparameters. **A6.1:** As you said, TOP$_{ll}$ is sensitive to hyperparameters, but **its actual negative effects are tiny**. - On the one hand, the learnable hyperparameters we used can effectively **self-adapt** to the data distribution and **avoid** the tedious manual parameter tuning. - On the other hand, the results under **all** hyperparameter configurations in our experiments are actually always **significantly higher** than the baseline (**16.0~23.9** v.s. 10.9 on Q3.1, **15.1~23.9** v.s. 10.9 on Table 3 in the paper), which further implies the effectiveness of our methods. ---- **Q6.2:** As shown in Figure 3 of the global rebuttal, the introduction of the similarity module removes many predictions of topology. But both results of (b) and (c) are far away from GT. I think that the existing metrics may not reflect the results well, and the filtering could lead to large variations. **A6.2:** We are very sorry that our **improper presentation** in Figure 3 gave you the misunderstanding. - To avoid the **visual confusion** caused by excessive topological connections, during visualization in Figure 4 of the main paper and its extended version Figure 3(b,c) of the Rebuttal PDF , only **the right lanes** (blue lines), **the incorrect lanes** (yellow lines) and **the incorrect topological connections** (red arrows) are drawn (as is description in the caption of Figure 4 of the main paper), while **the correct topological connections** are **omitted**, which is why the results of (b,c) look away from GT. - You can find that **all** the arrows in the figure are red (wrong), but it is **obviously impossible** that all connections are wrong (red), so please believe that we **deliberately visualized it this way** (although it looks stupid now), rather than **the result itself being bad**. In this perspective, what the figure shows is that the similarity module can further reduce **false predictions**, rather than **all predictions**. - Thank you very much for pointing out this issue. We will definitely add the correct predictions in the figure in the final version (using other color, e.g. green), but unfortunately, we **cannot upload** any pictures again during the discussion stage limited by Open Review system. If you still have doubts about the visualization results, you can also pay attention to **Figure 5** in the main paper again. We believe both the **Lane Topology** and the **Lane Graph** in it can definitely make people clearly understand the effect of our method. ---- **Q6.3:** The authors are not able to clearly explain the results of the DET. **A6.3:** It is indeed hard to provide a more clear explanation in a short time, however, compared with the little drop on DET$\_{t}$(**-1.4**), the **improvement in lane & topology** (**+13.0** TOP$_{ll}$, **+4.3** OLS, **+1.3** DET$_l$ [Topologic v.s. TopoNet]) is **more significant**. Besides, the latter is our **core research motivation and improvement goal**. Objectively speaking, compared with the **reproduced TopoNet**, our method is not inferior in terms of DET$\_t$ (**47.2** v.s. 46.9). It means that the issue lies in the reproduction of TopoNet (**46.9** v.s. 48.6) rather than our TopoLogic, which has also puzzled other researchers (e.g., the reproduction of TopoNet has lower DET$\_{t}$ (**44.5** v.s. 48.6) in Table 1 of SMERF paper). If there is still a chance, we will definitely explore this issue in depth and fix it. ---- **If you have any further questions or suggestions about our article, we'd love to discuss our content with you.**

Authorsrebuttal2024-08-14

Authors' Response to All Reviews

Thanks for the reviewers' careful reading of our paper, and giving valuable suggestions for us to improve the paper. We hope our response and clarifiction can ease some of your concern. If you have any further questions, we are more than happy to discuss with you if time permits. If you are satisfied after reading our response, we sincerely hope that you could increase the rating of our paper. Thank you.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC