RelTopo: Multi-Level Relational Modeling for Driving Scene Topology Reasoning

TL;DR

This work addresses the challenge of driving scene topology reasoning by identifying fragmentation between perception and reasoning and the underutilization of relational cues. It introduces RelTopo, a multi-level relational modeling framework that integrates relation-aware perception, geometry-informed L2L and cross-view L2T reasoning, and contrastive supervision to jointly optimize perception and topology inference. The approach leverages a geometry-biased self-attention, curve-guided cross-attention, and cross-view fusion to unify lane and traffic-element reasoning, with InfoNCE-based relational supervision to shape embedding spaces. Empirical results on OpenLane-V2 demonstrate state-of-the-art improvements across detection and topology metrics, underscoring the practical impact of embedding structural relations into end-to-end driving scene understanding.

Abstract

Accurate road topology reasoning is critical for autonomous driving, as it requires both perceiving road elements and understanding how lanes connect to each other (L2L) and to traffic elements (L2T). Existing methods often focus on either perception or L2L reasoning, leaving L2T underexplored and fall short of jointly optimizing perception and reasoning. Moreover, although topology prediction inherently involves relations, relational modeling itself is seldom incorporated into feature extraction or supervision. As humans naturally leverage contextual relationships to recognize road element and infer their connectivity, we posit that relational modeling can likewise benefit both perception and reasoning, and that these two tasks should be mutually enhancing. To this end, we propose RelTopo, a multi-level relational modeling approach that systematically integrates relational cues across three levels: 1) perception-level: a relation-aware lane detector with geometry-biased self-attention and curve-guided cross-attention enriches lane representations; 2) reasoning-level: relation-enhanced topology heads, including a geometry-enhanced L2L head and a cross-view L2T head, enhance topology inference via relational cues; and 3) supervision-level: a contrastive InfoNCE strategy regularizes relational embeddings. This design enables perception and reasoning to be optimized jointly. Extensive experiments on OpenLane-V2 demonstrate that RelTopo significantly improves both detection and topology reasoning, with gains of +3.1 in DET$_l$, +5.3 in TOP$_{ll}$, +4.9 in TOP$_{lt}$, and +4.4 overall in OLS, setting a new state-of-the-art. Code will be released.

Figures (6)

• Figure 1: Performance and visualization comparison between previous methods and ours. The top-left shows multi-view input images. Our approach, RelTopo, integrates relational modeling across multiple levels, strengthening both perception and topology reasoning. As shown in the bottom-left quantitative results on OpenLane-V2, RelTopo significantly outperforms prior methods across all metrics. On the right, qualitative comparisons demonstrate that RelTopo produces more accurate lane geometries and well-aligned connectivity than prior approaches.
• Figure 2:
• Figure 3: Illustration of our Bézier lane decoder layer, featuring our geometry-biased SA and curve-guided CA .
• Figure 4: Comparative visual results on OpenLane-V2 subsetA. The top row shows multi-view input images, the bottom row shows lane predictions. We show comparison between groundtruth, TopoMLP topomlp and ours. The blue box highlights misaligned connection point predictions from TopoMLP, and the green box shows the corresponding aligned predictions from our RelTopo . For clarity, zoomed-in views of selected regions are displayed at the top-right or bottom-right corners.
• Figure 5: Qualitative results comparison. The 1$^\text{st}$ row presents the multi-view input images, the 2$^\text{nd}$ row shows the predicted lane centerline results, and the 3$^\text{rd}$ row illustrates the predicted L2L topology results, with light blue arrowlines indicating directed connectivity between lanes. The final row depicts the front-view L2T topology predictions, where orange boxes highlight detected traffic elements (e.g., traffic lights and signs), mbevlane lines denote projected lanes, and blue lines in FV irepresent the pairing relationships between traffic elements and lanes. Each result row consists of three columns: the left column shows ground truth, the center column shows results from TopoMLP, and the right column presents our results. Two data samples are illustrated in this figure.