DB3D-L: Depth-aware BEV Feature Transformation for Accurate 3D Lane Detection

Yehao Liu, Xiaosu Xu, Zijian Wang, Yiqing Yao

3D lane detection is a cornerstone of autonomous driving, providing the spatial coordinates necessary for trajectory planning and lane keeping. However, extracting 3D information from a single 2D front-view (FV) camera is an "ill-posed" problem—the depth information is simply missing.

A new paper, DB3D-L, proposes a streamlined, depth-aware framework to bridge this gap without the computational overhead of heavy Transformers or the inaccuracies of "flat-ground" assumptions.

TL;DR

DB3D-L is an end-to-end 3D lane detection framework that integrates a dedicated Depth Net to transform front-view images into accurate Bird’s-Eye-View (BEV) features. By using efficient feature reduction and a spatial cross-attention fusion mechanism, it achieves state-of-the-art performance on synthetic and real-world datasets while remaining computationally efficient.

What Problem Does This Paper Address?

Most existing 3D lane detection methods fall into two camps, both with significant drawbacks:

  1. Geometry-based (IPM): These use Inverse Perspective Mapping to project pixels onto a ground plane. They rely heavily on the assumption that the road is perfectly flat. When the road is undulating or hilly, these models fail.
  2. Learning-based (LSS/Transformers): Methods like Lift-Splat-Shoot (LSS) or Vision Transformers (ViT) are powerful but computationally expensive. LSS often creates massive, sparse index matrices that are difficult to deploy, while Transformers require significant memory and data to converge.

The authors of DB3D-L sought a middle ground: a model that uses explicit depth estimation to guide the view transformation but does so using standard, efficient operators that are easy to deploy.

Methods

The DB3D-L architecture consists of three main modules:

1. D&F Net (Depth & Feature Network)

Instead of treating depth as an afterthought, the model uses a shared ResNet backbone to extract both FV features and depth information simultaneously. A Squeeze-and-Excitation (SE) layer is integrated to provide spatial attention, helping the network focus on crucial regions of the image.

2. FusionNet (The Core Innovation)

This module handles the transformation from FV to BEV through two sub-components:

  • Prime Feature Extraction (PFE): In a 2D image, distant objects often overlap in the vertical (height) dimension. However, in 3D space, lanes exist on a single plane. PFE uses column-wise max-pooling and convolutions to "compress" the height dimension, filtering out noise from elevated areas (like trees or buildings) and isolating the "prime" road features.
  • Depth Attention Module (DAT): This module uses the FV features to guide the depth features. It ensures that the depth probability distribution is spatially aligned with the visual features.
  • Hadamard Fusion: Finally, the BEV feature is built by taking the Hadamard product (element-wise multiplication) of the prime FV features and the prime depth features. This is significantly more efficient than the matrix multiplications used in Transformer-based cross-attention.

3. 3D Lane Detection Head

The resulting BEV map is divided into a grid. The head performs three tasks for each cell:

  1. Classification: Is there a lane here?
  2. Regression (X/Z): What is the lateral offset and the height (elevation) of the lane point within this cell?
  3. Clustering: An unsupervised method groups these points into individual lane instances.

Datasets

The model was evaluated on two primary benchmarks:

  • ApolloSim: A synthetic dataset featuring 10,000 samples across highway, urban, and residential scenes. It provides perfect ground truth for depth and 3D coordinates.
  • OpenLane v1.2: A massive real-world dataset containing 200,000 frames. It includes challenging scenarios like weather changes, intersections, and night driving.

Key Insights and Ideas

  • The "Height" Filter: One of the most clever aspects of DB3D-L is the explicit reduction of the height dimension. By recognizing that lanes are essentially 2D structures in a 3D world (on the ground), the authors simplify the feature space, making the fusion of depth and visual data much cleaner.
  • Standard Operators Only: Unlike many recent SOTA models that require custom CUDA kernels for LSS or complex Transformer blocks, DB3D-L uses standard convolutions and pooling. This makes it highly "deploy-friendly" for automotive hardware.
  • Depth Supervision is Vital: The ablation studies show that explicitly supervising the Depth Net (calculating a depth loss) improves the F1-score by over 2% and significantly reduces Z-error (height estimation).

Key Results

  • ApolloSim: In "Balanced Scenes," DB3D-L achieved an F1-score of 98.32%, outperforming established models like Gen-LaneNet and PersFormer. It also achieved the lowest lateral (X) error at near distances (0.0197m).
  • OpenLane: On real-world data, the model remained highly competitive, particularly excelling in near-distance X-error (0.2536m), which is critical for immediate lane-keeping maneuvers.
  • Efficiency: The ablation study on FusionNet showed that their approach reduced the parameter count compared to a basic fusion module while simultaneously boosting the F1-score from 94.92% to 98.15%.

Limitations and Future Work

While DB3D-L performs excellently in near-to-mid distances, the paper notes that errors (particularly in the Z/height dimension) naturally increase at far distances. This is a common limitation of monocular systems where pixel resolution for distant objects is low.

Future work could involve integrating temporal information (video) to further refine depth estimation or exploring how this lightweight architecture handles multi-camera setups for a full 360-degree BEV perception.

Summary updated

Abstract

3D Lane detection plays an important role in autonomous driving. Recent advances primarily build Birds-Eye-View (BEV) feature from front-view (FV) images to perceive 3D information of Lane more effectively. However, constructing accurate BEV information from FV image is limited due to the lacking of depth information, causing previous works often rely heavily on the assumption of a flat ground plane. Leveraging monocular depth estimation to assist in constructing BEV features is less constrained, but existing methods struggle to effectively integrate the two tasks. To address the above issue, in this paper, an accurate 3D lane detection method based on depth-aware BEV feature transtormation is proposed. In detail, an effective feature extraction module is designed, in which a Depth Net is integrated to obtain the vital depth information for 3D perception, thereby simplifying the complexity of view transformation. Subquently a feature reduce module is proposed to reduce height dimension of FV features and depth features, thereby enables effective fusion of crucial FV features and depth features. Then a fusion module is designed to build BEV feature from prime FV feature and depth information. The proposed method performs comparably with state-of-the-art methods on both synthetic Apollo, realistic OpenLane datasets.
BibTeX
@misc{liu2025db,
  author = {Yehao Liu and Xiaosu Xu and Zijian Wang and Yiqing Yao},
  title = {{DB3D-L: Depth-aware BEV Feature Transformation for Accurate 3D Lane Detection}},
  year = {2025},
  eprint = {2505.13266},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.13266},
}
Figure