Abstract
Autonomous driving without high-definition (HD) maps demands a higher level of active scene understanding. In this competition, the organizers provided multi-perspective camera images and standard-definition (SD) maps to explore the boundaries of scene reasoning capabilities. We found that most existing algorithms construct Bird's Eye View (BEV) features from these multi-perspective images and use multi-task heads to delineate road centerlines, boundary lines, pedestrian crossings, and other areas. However, these algorithms perform poorly at the far end of roads and struggle when the primary subject in the image is occluded. Therefore, in this competition, we not only used multi-perspective images as input but also incorporated SD maps to address this issue. We employed map encoder pre-training to enhance the network's geometric encoding capabilities and utilized YOLOX to improve traffic element detection precision. Additionally, for area detection, we innovatively introduced LDTR and auxiliary tasks to achieve higher precision. As a result, our final OLUS score is 0.58.
Introduction
The reliance on High-Definition (HD) maps presents a significant bottleneck for the scalability of autonomous driving systems due to high production costs and maintenance challenges. "Mapless" driving seeks to overcome this by enabling vehicles to reason about their environment in real-time. However, purely vision-based systems often struggle with long-range perception and occlusions. The MapVision team addresses these challenges by integrating Standard Definition (SD) maps—readily available from sources like OpenStreetMap—as a topological prior to guide the model's spatial reasoning.
Methodology
The MapVision framework is a multi-stage pipeline that fuses visual data with map-based geometric priors.
SD Map Encoding and Pre-training
Rather than treating SD maps as static images, the system encodes map lines using sine-cosine positional encoding and category information via one-hot encoding. These features are processed through a Transformer-based map encoder. To bolster the network's geometric understanding, the team implemented an AutoEncoder-based pre-training phase. By forcing the encoder to reconstruct map embeddings, the model develops a robust internal representation of road layouts before tackling the primary detection tasks.
BEV Construction and Architecture
The core architecture utilizes BEVFormer to project multi-view camera features into a unified Bird's Eye View (BEV) space. These visual features are then fused with the SD map features through a Map Cross-Attention mechanism.
Specialized Detection Heads
- Area Detection (LDTR): To improve the integrity of detected areas like pedestrian crossings, the team moved away from traditional keypoint-based methods. Instead, they adopted an "anchor-chain" approach inspired by LDTR, which treats areas as holistic instances.
- Traffic Element Detection: A heavily optimized YOLOX model was employed for 2D object detection, utilizing techniques such as Mosaic and Mixup data augmentation, a Path Aggregation Feature Pyramid Network (PAFPN), and Test Time Augmentation (TTA).
- Decoupled Training: The team observed that increasing the complexity of lane detection could negatively impact area detection. To mitigate this, they decoupled the training of the area head, allowing it to reach higher precision independently.
Experimental Results
The MapVision team evaluated their approach on the OpenLane-V2 benchmark. The integration of various components led to a final OpenLane Unified Score (OLUS) of 0.58.
Key findings from the ablation study demonstrate the incremental value of each architectural choice. Upgrading the backbone from ResNet-50 to InternImage-L provided a substantial boost in all metrics. Furthermore, the addition of the pre-trained map encoder and the LDTR head significantly improved lane and area detection scores.
The optimized YOLOX model achieved a traffic element detection score ($DET_t$) of 0.73, a significant improvement over the 0.55 baseline.
Conclusion
The MapVision report demonstrates that SD maps provide a powerful and cost-effective alternative to HD maps for autonomous navigation. By combining sophisticated BEV construction with geometric map priors and specialized detection heads, the team achieved state-of-the-art performance in the CVPR 2024 Mapless Driving Challenge. Future work will focus on resolving the trade-offs between detection precision and topological reasoning to further enhance scene understanding.
Summary updated