TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning

Summary

The paper presents a framework called TopoMLP for driving topology reasoning, which aims to accurately predict the relationship between lanes and traffic elements in autonomous driving scenarios. The key idea of TopoMLP is to leverage a simple yet effective approach based on MLP networks for topology reasoning. The authors also address the limitations of the topo metric of OpenLane V2 and provide an improved version. This topology reasoning algorithm achieves the state-of-the-art performance in the lane-lane and lane-traffic element relationship reasoning task on the OpenLane V2 datasets.

Strengths

1. The fundamental idea presented in the paper is solid, and the overall writing is of high quality. 2. The paper provides experimental results that effectively support the proposed method. 3. The authors enhance the query-based detectors by integrating additional reference boxes generated by YOLOv8. 4. The topology reasoning method employed is both simple and effective.

Weaknesses

1. One weakness of the proposed method is its inability to accurately represent complex shapes of map elements such as curbs using bezier curves. This limitation restricts the capability of the method for effective map learning and topology reasoning. 2. Additionally, while the authors have made improvements to the query-based traffic elements detection model, the box proposals generated are still outputs from a yolov8 model, which leaves some aspects of the work incomplete.

Questions

1. Regarding the scenarios depicted in Figure 3, where the intersection area on both the left and right sides is either invisible or not directly observable, the displayed results appear to be very promising. I would like to confirm whether these results were obtained by overfitting the dataset.

Rating

8: accept, good paper

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Summary

The paper presents a methodology named TopoMLP for topology reasoning in autonomous driving. Topology reasoning is essential for understanding road scenes and presenting drivable routes, requiring the detection of lane centerlines and traffic elements and reasoning their topology relationships, such as lane-lane and lane-traffic element topologies. The authors introduce a powerful 3D lane detector and an improved 2D traffic element detector, emphasizing the significance of detection performance on lane and traffic elements in influencing topology scores. They develop two simple MLP-based heads for topology generation, based on impressive detection performance. The proposed model, TopoMLP, achieves state-of-the-art performance on the OpenLane-V2 benchmark and ranks 1st in the 1st OpenLane Topology in Autonomous Driving Challenge. The authors also identify a loophole in the current topology metrics, and propose a correctness factor to enhance it.

Strengths

1. The authors conduct a comprehensive investigation into the significance of detection performance in topology reasoning, utilizing the OpenLane-V2 metrics. They highlight that with a detection score of 28.3, the topology score can only reach a maximum of 7.5, even if the topology reasoning results are identical to the ground truth. 2. In the above scenario, their model achieves a topology score of 7.2. This demonstrates the model's superior performance in topology reasoning compared to other models evaluated using the OpenLane-V2 metrics. 3. They propose using lane coordinate embedding and extra-supervision for interaction points to increasing LL topology reasoning performance. The LT topology can also benefited from view transformation embbeding and traffic element features. 4. They have established a strong baseline framework for topology reasoning for autonomous driving, and plan to make it open-source, laying a solid foundation for subsequent work. They also pointed out a flaw in an existing metric, which has contributed to the advancement and refinement of this field.

Weaknesses

1. Although the overall performance of TopoMLP is impressive, it primarily stems from the overall framework and training strategy, where the novelty being somewhat limited. 2. The several modules proposed in the paper to enhance performance are effective, but they are quite straightforward and intuitive, lacking more refined and intricate design. 3. Some parts of the manuscript are unclear and could be further improved; errors and suggestions for improvement are detailed later in the Questions.

Questions

1. TopoMLP has a higher lane-lane topology score (7.2) compared to the previous state-of-the-art TopoNet (4.1) under the similar detection score. However, even without your specific designs, the performance of TopoMLP remains strong in the ablation study Table 3(c) (6.9 and 6.5). Can you explain the main reason for the strong TOP_ll performance of your model compared to TopoNet? 2. Although the introduction of YOLOv8 can significantly improve performance, why not directly use YOLOv8's features for LT relationship inference, but use its detection results as proposals of a DETR? 3. Can you clarify why the "adjusted TOP" dropped in the "enhanced prediction"? Do the true positives and false positives in N_TP and N_FP refer to the matching in detection or topology? 4. Some errors in manuscript: - In Figure 2, the transformation matrix A is added to the Traffic Features. However, in Section 3.4, it is mentioned that it is "summed with the lane query embedding." - In Page 6, Section 4.2, the subscript of "The loss coefficient of lane-traffic topology loss" should be "lt" instead of "ll".

Rating

6: marginally above the acceptance threshold

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Summary

The paper proposes a simple and strong method called TopoMLP, on the openlane-v2 benchmark for topology reasoning in the driving scenarios. Authors claim that the key insight is to improve the detection part in order to secure a steady improvement of the topology score. This is also the winner solution for the Openlane-v2 challenge.

Strengths

In general the paper carries some good merits, including: + the paper is easy to follow. The contributions are very clear: good detection increases the topology reasoning. Based on the detection module, using two simple MLP heads for topology generation could win among other approaches. + this work derives from the participation of this year's Autonmous Driving Challenge. This is the first entry solution, as shown in Table 1 and Table 2. Authors find a problematic on the evalution metric of OLS score. Figure 5 is very illustrative. Using the rectified evaluation, the loophole could be remedied.

Weaknesses

The paper has some obvious shortcomings as follows. - Motivation / technical novelty seems to be limitted. Following the second paragrah in the introduction, it is empirically found that topology performance is improved with stronger detection. Any insight behind this? There seems not too many discussions on it. - Some ablations and experiments need to be discussed. Overall this work falls into the borderline category without too many novelty or insight. However, given (a) the topology reasoning for lane detection is an emerging topic and there are few work in this direction, (b) the paper achieves the best result on the leaderboard of the Challenge, promised reproducible research to the community, the reviewer would slightly lean to encourage the manuscript to be polished in great extent.

Questions

1. I see the pipeline resembles to some extent with TopoNet method, where two branches are capable of recognizing the traffic lights and lane topoploy respectively. Using two simple MLP for the head is the core contribution I suppose. Is there any ablations to verify that using some complicated design (like a Transformer head) could "downgrade" the OLS score? 2. Section 4.6 is a good contribution to the open-sourced community. Do authors communicate with organizers of the Challenge to reconsider the metric and maybe the leaderboard would be refreshed? On further consideration, the current OLS is a general metric. Is it good to design a separate metric to evaluate the topology performance at intersections (which is crucial I suppose)? 3. This work generates the centerline of the lane, which is not the physical representation of the road scene. Is it a good formulation? On a higher level, I think the current form of Openlane-v2 tackles only partial challenges in the static road recognition. How to utilize the topoloy to faciliate downstream task? This could potentially trigger more future research in related domains. ---- Minor 1. Topo in Figure 2. "Topoogy" 2. G^LL in Equation 3 is the operation of Topology in Figure 2? I suggest to align the concept with equation and the figure. 3. How to balance the different weights in Equation 5 of different branches? Are there some ablations or just equally distributed?

Rating

6: marginally above the acceptance threshold

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Summary

The authors introduces TopoMLP driving topology reasoning using vehicle-mounted sensors. The model starts with two strong decoder to detect the lanes and the traffic elements. For the lanes, the model uses Bezier curves as lane queries for 3D centerline position embedding. For the traffic elements, it uses proposals from YOLOv8 and the corresponding multi-scale features, and feed them to a series of deformable decoder layers. These information then are passed to two MLP network reasoners to perform the final lane-lane and lane-traffic-element topology reasoning. Both of these modules apply pair-wise representations for their respective inputs. However, Lane-Traffic reasoning also takes in camera intrinsic and extrinsic information. The loss functions uses focal loss throughout, for classification for both lane and traffic detection, and lane-lane and lane-traffic topology. In addition, lane detection loss also includes L1 loss for lane regression and lane-lane topology, while traffic detection loss also includes GIoU loss for bounding box regression. TopoMLP is tested on OpenLane-V2 benchmark, 2000 scenes in total, against a number of STOA techniques. The model is able to surpass the rest of the techniques, by showing a much better topological reasoning capability, while achieving a decent detection accuracy. The ablation study shows that the value of using YOLOv8 object detection, explicit lane coordinates, and injecting transformation matrix. Finally, the paper proposes an enhanced TOP score incorporating correctness factor, which allows for more accurate precision.

Strengths

The proposed model improves the STOA on OpenLane dataset by refining its various components. The proposed enhanced TOP score also shed lights to a more accurate evaluation.

Weaknesses

It is hard to make sense of why the TOP numbers are quite low, especially lane-lane topology. It would be good to add what is an acceptable number for robust driving. There is not a lot of discussion of what is still not solvable, to allow readers to understand the shortcomings of the methods. This would add discussions on what is the next steps to take.

Questions

Would you be able to ground the TOP numbers with what is required for autonomous driving. Also discussions on what is not yet solvable. Typos, Grammar, etc.: pg. 4, figure 2 caption: “Topoogy” —> “Topology”

Rating

6: marginally above the acceptance threshold

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Ethics concerns

None.