Anchor3DLane++: 3D Lane Detection via Sample-Adaptive Sparse 3D Anchor Regression

TL;DR

This work tackles monocular 3D lane detection by moving away from BEV-based methods and IPM limitations, proposing Anchor3DLane++ which predicts 3D lanes directly from front-view features using 3D lane anchors. A key contribution is the Prototype-based Adaptive Anchor Generation (PAAG), which builds sample-adaptive sparse anchors from learned metas, paired with an online Equal-Width (EW) loss to regularize parallel lane geometry. The framework supports cross-layer iterative refinement and optional camera-LiDAR fusion, achieving state-of-the-art results on OpenLane, ApolloSim, and ONCE-3DLanes while maintaining real-time performance. Overall, this BEV-free, anchor-based approach provides accurate 3D lane predictions with efficient sampling and strong generalization across datasets and sensor modalities.

Abstract

In this paper, we focus on the challenging task of monocular 3D lane detection. Previous methods typically adopt inverse perspective mapping (IPM) to transform the Front-Viewed (FV) images or features into the Bird-Eye-Viewed (BEV) space for lane detection. However, IPM's dependence on flat ground assumption and context information loss in BEV representations lead to inaccurate 3D information estimation. Though efforts have been made to bypass BEV and directly predict 3D lanes from FV representations, their performances still fall behind BEV-based methods due to a lack of structured modeling of 3D lanes. In this paper, we propose a novel BEV-free method named Anchor3DLane++ which defines 3D lane anchors as structural representations and makes predictions directly from FV features. We also design a Prototype-based Adaptive Anchor Generation (PAAG) module to generate sample-adaptive sparse 3D anchors dynamically. In addition, an Equal-Width (EW) loss is developed to leverage the parallel property of lanes for regularization. Furthermore, camera-LiDAR fusion is also explored based on Anchor3DLane++ to leverage complementary information. Extensive experiments on three popular 3D lane detection benchmarks show that our Anchor3DLane++ outperforms previous state-of-the-art methods. Code is available at: https://github.com/tusen-ai/Anchor3DLane.

Figures (7)

• Figure 1: (a) BEV-based methods, which perform lane detection in the warped BEV images or features. (b) Non-BEV method, which projects 2D lane predictions back to 3D space with estimated depth. (c) Our Anchor3DLane++ projects 3D anchors into FV space to sample features for direct 3D prediction.
• Figure 2: The overall architecture of Anchor3DLane++. (a) Pipeline of Anchor3DLane++. Proposals output from the previous stage are used as the new anchors for the next stage. (b) The $r$-th iterative stage of anchor projection and 3D lane prediction. 3D anchors or 3D proposals from $(r+1)$-th stage are projected to sample their features from $\mathbf{F}^r$ using camera parameters. A classification head and a regression head are applied after a self-attention layer to make prediction.
• Figure 3: Illustration of 3D anchor and 3D lane in the ground coordinate system.
• Figure 4: Illustration of Prototype-based Adaptive Anchor Generation.
• Figure 5: Comparison of F1 score vs. throughput for different methods.