HD Maps are Lane Detection Generalizers: A Novel Generative Framework for Single-Source Domain Generalization

TL;DR

The paper tackles lane-detection generalization under domain shifts using a single-source SSDG setup with HD Maps as the data source. It introduces a diversity-aware coreset framework that (i) extracts lane masks from HD Maps, (ii) generates lane-conditioned images via SIS, and (iii) selects a diverse core set by optimizing lane-structure and surrounding diversity ($K_ ext{l}$, $K_ ext{i}$). Training on this core set yields improvements over a domain-adaptation baseline (MLDA) on CULane by $+3.01$ percentage points, with competitive gains on TUSimple, all without target-domain data. This demonstrates that HD Maps can serve as scalable, label-efficient sources for SSDG in lane detection, enabling robust deployment across unseen environments by synthesizing and curating diverse, label-consistent training data.

Abstract

Lane detection is a vital task for vehicles to navigate and localize their position on the road. To ensure reliable driving, lane detection models must have robust generalization performance in various road environments. However, despite the advanced performance in the trained domain, their generalization performance still falls short of expectations due to the domain discrepancy. To bridge this gap, we propose a novel generative framework using HD Maps for Single-Source Domain Generalization (SSDG) in lane detection. We first generate numerous front-view images from lane markings of HD Maps. Next, we strategically select a core subset among the generated images using (i) lane structure and (ii) road surrounding criteria to maximize their diversity. In the end, utilizing this core set, we train lane detection models to boost their generalization performance. We validate that our generative framework from HD Maps outperforms the Domain Adaptation model MLDA with +3.01%p accuracy improvement, even though we do not access the target domain images.

Figures (5)

• Figure 1: The overall framework. Our framework adopts HD Maps as a single-source domain to build robust lane detection models. We first extract numerous lane masks from HD Maps with surrounding images. Then, we train SIS (Semantic Image Synthesis) generative models to synthesize lane-conditioned images with diverse surroundings. We select a core set among synthesized images with two criteria to maximize its diversity: (i) lane structure and (ii) surroundings. Then, we finally train the lane detection model and obtain enhanced generalization performance.
• Figure 2: Examples of generated dataset. We utilized OASIS sushko2020you as our generative model to visualize results.
• Figure 3: Performance degradation by domain discrepancy. (a) Example images of source and target domains and (b) performance of the state-of-the-art lane detection model wang2022keypoint on the target domain, still revealing significant challenges due to domain differences.
• Figure 4: Domain discrepancy in lane detection. We categorized the cause of domain discrepancy in lane detection. (a) The difference in the number of lanes and curvatures shows a clear discrepancy in lane structures. (b) Judging from the t-SNE tsne analysis, the contextual difference of surroundings is also confirmed.
• Figure 5: Effect of the number of selected data. We compare the performance of randomly selecting 10 data from 100 images generated from a single lane label mask and selecting 2, 5, and 10 data using the proposed diversity-aware image selection method.