Enhancing the Effective Receptive Field of a U-Net-Based CNN for Road Marking Segmentation From Mechanical Spinning Mobile LiDAR-Derived Imagery

Abstract

This study presents a framework for segmenting road markings from sparse point cloud-derived imagery, addressing feature sparsity. Accurate segmentation in this context requires capturing both sparse features and broader spatial relationships. In deep learning, the effective receptive field (ERF) denotes the uneven input region influencing an output, and when restricted, networks struggle to capture sparse features. To address this, we propose two complementary modules: shape-aligned convolutions (SAC), which enhance spatial context, and the learnable feature resizer (LFR), which preserves fine details across layers. These modules are integrated into a unified architecture that captures both fine-grained features and broad spatial context, effectively expanding the ERF. To support this task, we collected two new datasets using a downward-tilted Velodyne-16 LiDAR sensor: the Minatomirai (MM) dataset in Japan and the University of the Philippines Diliman (UPD) dataset, totaling 15 100 paired intensity images with annotated labels. Evaluations against U-Net variants demonstrate superior median F1-scores, achieving 87.3% for mean classification, an 84% lower parameter count compared to the strongest competitor, and competitive training and inference speeds. Taken together, these findings highlight four key contributions: 1) SAC for broader spatial contextual relationships, 2) LFR for improved feature preservation, 3) their integration into an efficient segmentation framework, and 4) the introduction of two sparse point cloud-derived road-marking segmentation datasets, which are publicly available at this link.