TopoStreamer: Temporal Lane Segment Topology Reasoning in Autonomous Driving
Yiming Yang, Yueru Luo, Bingkun He, Hongbin Lin, Suzhong Fu, Chao Zheng, Zhipeng Cao, Erlong Li, Chao Yan, Shuguang Cui, Zhen Li
TL;DR
TopoStreamer is an end-to-end temporal perception model that improves lane segment topology reasoning in autonomous driving by enforcing temporal consistency and enhancing positional encoding, leading to better road network reconstruction.
Contribution
It introduces streaming attribute constraints, dynamic lane boundary positional encoding, and lane segment denoising to address limitations in existing methods.
Findings
Achieves +3.0% mAP in lane segment perception on OpenLane-V2
Achieves +1.7% OLS in centerline perception on OpenLane-V2
Demonstrates significant improvements over state-of-the-art methods
Abstract
Lane segment topology reasoning constructs a comprehensive road network by capturing the topological relationships between lane segments and their semantic types. This enables end-to-end autonomous driving systems to perform road-dependent maneuvers such as turning and lane changing. However, the limitations in consistent positional embedding and temporal multiple attribute learning in existing methods hinder accurate roadnet reconstruction. To address these issues, we propose TopoStreamer, an end-to-end temporal perception model for lane segment topology reasoning. Specifically, TopoStreamer introduces three key improvements: streaming attribute constraints, dynamic lane boundary positional encoding, and lane segment denoising. The streaming attribute constraints enforce temporal consistency in both centerline and boundary coordinates, along with their classifications. Meanwhile,…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The authors rightly choose to incorporate temporal information into the lane segment topology reasoning task, as it helps address missed detections caused by occlusions and high-speed motion—a critical issue in topology reasoning. 2. The two challenges highlighted by the authors, consistent positional embedding and temporal multiple attribute learning for lane segments, are well-motivated and meaningful. Moreover, their proposal of a new metric to evaluate lane boundary classification accura
1. I think the authors lack a detailed description of Section 3.4, LANE SEGMENT DENOISING. Why is denoising needed for predicting the topological relationships of these fine-grained lane segments? I hope the authors can further explain the motivation behind denoising, preferably with some simple examples. 2. Since introducing temporal information incurs significant computation, I hope the authors can provide the training and inference times.
- The paper proposes a temporal perception model that explicitly addresses temporal consistency and positional embedding issues in lane topology reasoning. - The paper includes comparisons with multiple strong baselines and extensive quantitative ablations, demonstrating solid engineering effort. The proposed method achieves state-of-the-art (SOTA) results in both lane segment and centerline perception tasks.
- The main contributions — temporal propagation, positional encoding refinement, and query denoising — are all direct extensions of existing methods.The proposed modules are minor architectural tweaks rather than conceptual advances. - The paper lacks a detailed analysis of computational complexity (e.g., FLOPs, parameter count) and its comparison with baseline methods.
1. Introducing temporaling modeling into the lane segment perception task is valuable for stable map construction. 2. Experimental results show that on the OpenLane-V2 dataset, TopoStreamer achieves a 3.0% mAP improvement in lane segment perception and a 1.7% OLS improvement in centerline perception compared to the previous state-of-the-art (SOTA)
1. Many parts of this work claimed as core contributions have similar spirits to other autonomous driving works like StreamMapNet, and etc. Overall, this paper mainly solves several problems when such temporal modeling is applied to this specific lane segmentation prediction task. 2. Streaming Attribute Constraints seems to be an auxiliary loss that adds direct supervision to the results of stream queries. Compared to the previous transformation loss, it seems to have a similar effect but is im
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAutonomous Vehicle Technology and Safety · Advanced Neural Network Applications · Adversarial Robustness in Machine Learning