Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation

Predicting and constructing road geometric information (e.g., lane lines, road markers) is a crucial task for safe autonomous driving, while such static map elements can be repeatedly occluded by various dynamic objects on the road. Recent studies have shown significantly improved vectorized high-definition (HD) map construction performance, but there has been insufficient investigation of temporal information across adjacent input frames (i.e., clips), which may lead to inconsistent and suboptimal prediction results. To tackle this, we introduce a novel paradigm of clip-level vectorized HD map construction, MapUnveiler, which explicitly unveils the occluded map elements within a clip input by relating dense image representations with efficient clip tokens. Additionally, MapUnveiler associates inter-clip information through clip token propagation, effectively utilizing long-term temporal map information. MapUnveiler runs efficiently with the proposed clip-level pipeline by avoiding redundant computation with temporal stride while building a global map relationship. Our extensive experiments demonstrate that MapUnveiler achieves state-of-the-art performance on both the nuScenes and Argoverse2 benchmark datasets. We also showcase that MapUnveiler significantly outperforms state-of-the-art approaches in a challenging setting, achieving +10.7% mAP improvement in heavily occluded driving road scenes. The project page can be found at https://mapunveiler.github.io.

Paper

Similar papers

Peer review

Reviewer Hway5/10 · confidence 5/52024-07-09

Summary

This paper aims to improve vectorized HD map construction for autonomous driving. Inspired by the global feature association in traditional offline HD mapping, the proposed MapUnveiler processes input frames in a clip-based manner and hopes to resolve occlusions using information from previous frames. Built up MapTRv2, MapUnveiler introduces clip tokens together with the Inter-clip and Intra-clip Unveiler modules to update the map queries with temporal information. Experiments on nuScenes and Argoverse2 datasets demonstrate the superior performance of the proposed method, especially on highly-occluded scenes.

Strengths

1. The idea of incorporating and aggregating clip-level information for online vectorized HD mapping is reasonable and is more akin to how humans drive. The proposed method has more thoughtful designs than early works such as StreamMapNet to better handle occlusions and incorporate long-range information. 2. The proposed MapUnveiler obtains state-of-the-art results in various experimental settings. The improvements over previous methods are especially prominent in the large 100mx50m setting and the highly-occluded scenes collected by the authors. 3. Extensive ablation studies enumerate the choices of almost all hyper-parameters or model components, which helps better understand and break down each element's contributions.

Weaknesses

1. The clarity of the method description is poor, making it very hard to thoroughly understand the proposed architecture. Details are discussed below: - The method explanation is not self-contained: i) The Inter-clip Unveiler section refers to the TTM and directly skips all details. There is no information at all about how is the compact memory token generated from the denser map queries; ii) The "loss" section refers to MapTRv2 and again skips all details. The authors should not assume the general audience to be aware of the concrete details of TTM and MapTRv2. The core formulation of these components should be elaborated with texts or equations, while full details can go to the appendix. - The definitions of the temporal window T and the stride S are unclear. Based on the text descriptions and the common definition of stride, my understanding of "T=3 and S =2" is that "each clip has 3 frames, and every two consecutive frames have a temporal gap of 1." However, the symbols in L177-178 seem to suggest other meanings of T and S. - The description of the inference mechanism is also vague. Is the MapUnveiler executed per frame or per clip? Figure 2 seems to suggest the per-clip inference where the predictions of T frames are obtained together. If this is the case, does it hurt the actual response frequency? In short, Section 3 of the paper lacks significant details, and I cannot properly understand MapUnveiler's exact formulation. Given that the authors answer "No" to Question 5 of the Checklist, I have to raise concerns about the paper's reproducibility. 2. There is no detail on how the pre-training and fine-tuning are conducted. Do you initialize the MapNet by training MapTRv2? If this is the case, how are the training epochs split for the MapNet pre-training and the end-to-end MapUnveiler fine-tuning? If the 24/6 epochs for nuScenes/Argo2 are only for the fine-tuning stage, then the comparisons in the main table are unfair, as other methods in the table have not fully converged. 3. The main comparison results are incomplete. Most previous papers provide the nuScenes results of both short and long training schedules, but the main table only presents short-schedule results. Considering the last question about the pre-training and fine-tuning, the authors should complement the table with long-schedule results to show that MapUnveiler can obtain consistent performance boosts when all the methods are fully converged. This concern is backed up by the fact that MapUnveiler's improvement is much smaller on Argo2 compared to nuScenes -- based on my empirical experience, previous methods like MapTRv2 and its followups converge faster on Argo2, and training for 6 epochs is close to convergence. This probably suggests that the large performance gaps on nuScenes come from unfair training settings. 4. Your interpretation of StreamMapNet and SQD-MapNet's Argo2 training epochs is wrong. These two methods employ a different frame sampling strategy at training time compared to MapTRv2, but their effective number of training samples is the same as MapTRv2. Therefore, the claim about the "longer training schedules" in the main table's caption is misleading. 5. The venues in the main table are not accurate. HIMap[49] and MGMap[24] are accepted by CVPR2024, and the information was already available at the time of NeurIPS submission. Furthermore, a recent HD map construction method, MapTracker[A], also studies temporal modeling and should be very relevant, but it is missing in the discussion and related works. [A] MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping, arXiv:2403.15951

Questions

The paper studies an important problem (temporal information association) in online HD map construction and proposes a reasonable method. However, the poor clarity and the potentially incomplete/unfair comparison results raise serious concerns about the paper's quality and reproducibility. My current rating is reject, and I will consider changing the score if the main weaknesses are properly addressed.

Rating

5

Confidence

5

Soundness

3

Presentation

1

Contribution

2

Limitations

The limitations and broader impacts are adequately discussed in the paper.

Reviewer Hway2024-08-09

Thanks for the detailed response. The added details clarify the architectural design, especially for the clip-based inference mechanism and how $T$ and $S$ work. I believe these will also help other readers better understand your approach. I appreciate the new results in the general response, which resolve my concern about the fairness of experiments. However, please make sure the results in G4 also follow a fair setting, or you should at least indicate the difference in training schedules in the table. Overclaiming the performance boost with an unfair setting under the hood will hurt the community. With the above concerns addressed, I will increase my score from 3 to 5 (borderline accept). Reasons for not giving a higher score are mainly two-fold: 1) the clip-wise inference is a bit unnatural, making the output quality temporally inconsistent; 2) the initial writing clarity is very problematic, requiring a massive revision to improve the paper's quality. If the paper is accepted, I hope the authors can carefully improve the clarity of the method description (W1), fix the discussion/reference of related works (W5), provide fair experimental results and clearly indicate the difference in training settings (W2,3).

Authorsrebuttal2024-08-10

Thank you!

We would like to express our sincere gratitude for your positive feedback and for raising your rating of our work from 3 to 5. We especially appreciate your thorough and professional comments, which have helped us improve the fairness of our experimental setup and enhance our presentation. As you suggested, we will revise the relevant results presented in the comparison tables (Table 1 and G4) with more fair settings (12/3 epochs fine-tuning) in the final version. Additionally, we will revise the final version by incorporating all comments from the rebuttal, including clarifications of our approach (W1), discussions of the suggested related works (W5), and clarifications of the training settings (W2, W3). Best regards, \ The Authors

Reviewer gpqV7/10 · confidence 3/52024-07-13

Summary

The authors propose a new approach for constructing vectorized high-definition maps that exploits temporal information across adjacent input frames. The model, which they call MapUnveiler, operates at the clip-level and consists of an intra-clip unveiler which generates vectorized maps for T frames and an inter-clip unveiler which uses a memory module to aggregate information between clips. The authors present results on two standard benchmarks, vectorized HD map construction benchmarks (nuScenes and Argoverse2) and demonstrate the model’s superior quantitive performance to several previously proposed approaches. They also show several qualitative examples of how MapUnveiler can better handle occlusions in the input images.

Strengths

- The paper is well-written and contextualized well within prior work. - The methodology is novel and well-motivated. - The results are strong on the two tested datasets, both quantitatively and qualitatively. - Many different analyses and ablations were included to justify the design decisions used within MapUnveiler and show its strengths.

Weaknesses

1. The methods is dense and a bit hard to read. The architecture figures help but are also a bit difficult to parse through. It would be helpful to try to weave more intuition into the text. 2. Claiming "-9.8%" is significant but "-6.0%" is comparable in the robustness to occlusion section seems a bit arbitrary (and potentially overstating MapUnveiler's performance, as a 6% drop is still considerable). I suggest the authors rephrase this sentence (and address similar claims in the paper). There are several typos throughout the paper. I have enumerated some here, but encourage the authors to do a detailed proofread: - 127: With there - 129: mapnet -> MapNet - 161: bev -> BEV - 167 parenthesis - 192 backwards parenthesis - 294: In addition, if we choose too short

Questions

1. Have the authors tried quantized models to reduce GPU memory? It could be interesting to see if the gains from larger window sizes outweigh the losses from quantization. 2. The model still seems to struggle with some occlusions (a 6% drop from the standard split). Why do the authors think that is? Are these just very difficult cases or issues with the model? 3. The one limitation that was discussed seems like it can be tested. How does randomly dropped intermediate frames affect model performance?

Rating

7

Confidence

3

Soundness

4

Presentation

3

Contribution

4

Limitations

Only one limitation is included. I encourage the authors to think through other potential limitations.

Authorsrebuttal2024-08-09

Thank you!

We sincerely appreciate your efforts in reviewing our responses. We are delighted to receive your positive feedback and supports for our work. Thanks,\ The Authors

Reviewer xabw6/10 · confidence 5/52024-07-14

Summary

This work presents a method called MapUnveiler, which aims to improve the construction of vectorized HD maps for autonomous driving. MapUnveiler uses a novel clip-level pipeline to unveil occluded map elements by relating dense image representations with efficient clip tokens and propagating inter-clip information. This approach leverages temporal information across adjacent input frames, addressing the limitations of single-frame and streaming inference methods. The model achieves state-of-the-art performance on the nuScenes and Argoverse2 benchmark datasets, demonstrating promising improvements in challenging scenarios with longer perception ranges and heavy occlusions.

Strengths

1. The introduction of a clip-level pipeline for vectorized HD map construction effectively addresses occlusion issues and leverages temporal information across multiple frames. 2. The method utilizes clip tokens to propagate map information efficiently, reducing redundant computations and enhancing prediction consistency. 3. Extensive experiments demonstrate that MapUnveiler achieves state-of-the-art performance on nuScenes and Argoverse2 benchmarks, particularly in challenging scenarios.

Weaknesses

1. The community has noticed a severe data leakage issue with utilizing nuScenes and Argoverse2 datasets for online mapping evaluation {1, 2}, as these datasets are not intentionally built for online mapping. It might also be necessary to validate the proposed method on geo-disjoint training and validation sets. 2. It would be good to see the analysis of added model compacity due to the introduction of the proposed intra-clip unveiler and inter-clip unveiler. 3. It seems the proposed intra-clip unveiler and inter-clip unveiler are adaptable to any single-frame inference online mapping methods. It would be good to validate the effectiveness of the proposed modules on other baseline methods. 4. The authors are encouraged to investigate the consistency of estimated HD maps across frames of the proposed method compared to existing methods with "inconsistent and suboptimal prediction results" (mentioned in Line 7). {1} Augmenting Lane Perception and Topology Understanding with Standard Definition Navigation Maps. {2} Localization Is All You Evaluate: Data Leakage in Online Mapping Datasets and How to Fix It.

Questions

1. What do the map queries stand for? Can they be transferred directly to vectorized HD maps? 2. Is the map decoder adopted from MapTRv2? 3. Are map tokens generated from the intra-clip unveiler the refined version of map queries?

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The limitation of dependency on temporally consecutive frames is discussed.

Reviewer JH5U5/10 · confidence 4/52024-07-15

Summary

This paper proposes a clip-based vectorized HD map construction paradigm for the processing of long temporal sequence, in which occluded map elements are unveiled explicitly by efficient clip tokens. Through clip token propagation, MapUnveiler achieves effective utilization of long-term temporal map information by associating inter-clip information, in which clip tokens are propagated rather than dense BEV features. Experiments demonstrate that MapUnveiler boosts the performance on public benchmark datasets, also for more challenging setting like long-range perception and heavily occluded driving scenes.

Strengths

1. This paper is well-written and easy-to-follow. Figures clearly conveys the intended message. 2. “Unveiling the hidden” and clip token propagation are reasonable and effective strategy for static Map element detection, which is practical and alleviates the problem to some extent. 3. The proposed method demonstrates strong performance on benchmark dataset, comprehensive experiments and ablation studies justify the model design.

Weaknesses

1. As mentioned at line 227, this work is built on pretrained frame-level MapTRV2 and fine-tuned, thus the comparison can be unfair. Results without pretraining are required to verify your effectiveness. 2. At line 53 and BEV Updater in line 151, for occluded features, how to select the tokens that are visible in certain frames? Seems tokens within the temporal window are fully utilized for BEV update by cross attention, how to determine whether these tokens contain unblocked information? More explanations are required.

Questions

What is the experiment result for geometric-based dataset split as mentioned in [1] and [2]? Besides, what is the additional computing costs considering the injection of temporal clip token? [1] Yuan, Tianyuan, et al. "Streammapnet: Streaming mapping network for vectorized online hd map construction." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024. [2] Lilja, Adam, et al. "Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes. The authors mentioned the weakness of their approach on the corrupted input.

Reviewer gpqV2024-08-09

Thank you for your substantial efforts in the overall rebuttal and in response to my specific comments. The changes you have proposed and the explanations of the results help considerably. Thanks for explaining the difficulties with implementing a quantized model. I still think it would be valuable to see so it would be great if you could try to get it working. That being said, I don't think it's critical to the paper. I have read the other reviews + rebuttal, and maintain my score. However, I encourage the other reviewers to reconsider their assigned scores after the authors' rebuttal.

Reviewer xabw2024-08-11

Thank you for the thorough and thoughtful rebuttal. The authors have successfully addressed my concerns, so I am increasing my rating from 5 to 6.

Authorsrebuttal2024-08-12

Thank You!

We are so pleased to have received your positive feedback and raised score (5 → 6). We deeply appreciate your efforts in reviewing our responses. Best regards,\ The Authors

Authorsrebuttal2024-08-14

Thanks again for your review

Dear Reviewer JH5U, We greatly appreciate your valuable efforts and professional feedback, which have indeed improved the quality of the final version of our manuscript. We have provided answers to your remaining concerns above, and it would be great to hear your feedback on our rebuttal so that we can further improve the final version. Although the authors-reviewers discussion period is nearing its end, we are fully prepared to address any further questions you may have. Best regards,\ The Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC