A Comprehensive Satellite Imagery Dataset for Road Marking Detection

High-definition (HD) maps play a critical role in autonomous driving by providing precise information about the road environment. To advance the automation of HD map generation, we introduce SEED-MAP, a large-scale, high-resolution satellite imagery dataset that delivers high-quality annotations for road markings. By providing the foundational data to train road marking recognition models, our dataset directly addresses a key prerequisite for automating the HD map construction pipeline. SEED-MAP is the first large-scale dataset to offer detailed annotations for both lane markings and diverse road symbols, such as crosswalks and arrows. The dataset offers dual-level annotations that capture both the visual appearance and functional meaning of each lane marking. The annotations include pixel-level and geographic coordinates, supporting both machine learning and real-world navigation. To demonstrate the dataset's utility, we established performance benchmarks for state-of-the-art object detection and semantic segmentation models. SEED-MAP is publicly available, providing an accessible resource for researchers to develop and benchmark perceptual models for satellite imagery. By offering a unique combination of comprehensive annotations and geographic precision, SEED-MAP aims to drive innovation in HD map construction and satellite imagery analysis.