Abstract
Understanding road topology—the intricate relationships between lanes and traffic lights—is crucial for autonomous navigation in complex urban environments. TopoMLP demonstrates that the primary bottleneck in topology reasoning is not the complexity of the relational model, but the quality of the underlying object detection. By pairing robust 3D lane and 2D traffic detectors with simple Multi-Layer Perceptron (MLP) heads, TopoMLP achieves state-of-the-art performance on the OpenLane-V2 benchmark, winning the 1st OpenLane Topology Challenge.
Introduction
For an autonomous vehicle to navigate a complex intersection, it must perceive more than just isolated objects; it must understand the scene's topology. This involves reasoning about connectivity: which lanes lead into others (lane-lane topology) and which traffic lights govern specific lanes (lane-traffic topology). While existing research often focuses on sophisticated graph neural networks (GNNs) to model these relationships, TopoMLP identifies a fundamental "perception-first" bottleneck. If a detector fails to identify a lane or a traffic light initially, no reasoning mechanism can recover the missing connection. TopoMLP quantifies this dependency and proposes a simplified pipeline that prioritizes strong detection.
Methods
TopoMLP follows a "first-detect-then-reason" philosophy, utilizing a modular architecture with two primary detection branches followed by lightweight reasoning heads.
- 3D Lane Detector: Inspired by PETR, this branch represents road centerlines as smooth Bézier curves. A transformer-based decoder predicts 3D control points, providing a more geometrically consistent representation than traditional point-cloud methods.
- 2D Traffic Element Detector: Utilizing a Deformable DETR head, this branch identifies traffic lights and signs. To improve detection of small, distant objects, an optional YOLOv8 detector can be integrated to provide high-quality proposals.
- MLP Topology Heads: Instead of complex GNNs, TopoMLP uses simple MLPs. It concatenates query embeddings from the detectors with spatial position embeddings. The MLP then performs binary classification to determine if a relationship exists between pairs of elements.
- Metric Correction: The authors proposed an adjusted TOP metric for OpenLane-V2 that incorporates a "correctness factor" to prevent models from gaming scores through high-confidence false positives.
Key Insights and Results
The core finding of TopoMLP is that topology reasoning is a detection-limited task. An ablation study revealed that replacing predicted topology with Ground Truth (GT) connections while keeping detectors frozen yielded minimal gains (e.g., ~0.5% for lane-lane topology). Conversely, simply upgrading the detection backbone resulted in significant performance leaps.
Performance Highlights:
- State-of-the-Art: On OpenLane-V2 subset_A, TopoMLP achieved 41.2% OLS (OpenLane-V2 Score) with a ResNet-50 backbone, significantly outperforming previous methods like TopoNet (35.6%).
- Challenge Success: The pipeline secured 1st place in the 1st OpenLane Topology in Autonomous Driving Challenge.
- Scalability: Performance improved consistently as the backbone strength increased (from ResNet-50 to Swin-B), validating the "detect-then-reason" strategy.
Conclusion
TopoMLP proves that simple MLP-based reasoning is sufficient when supported by high-quality feature embeddings from strong detectors. While highly effective, the model's modular dependency means it remains vulnerable to detection failures in adverse conditions where markings are invisible. Future work may involve incorporating temporal context from video sequences to maintain consistency and exploring end-to-end refinement loops where reasoning informs the detection process.
Summary updated