Bai, Zenan (Xi 'an Jiaotong University), Yonghong, Song (Xi'an Jiaotong University), Meng, Zeyu (Xi'an Jiaotong University), Xiaomeng, Wu (Xi'an Jiaotong University)

LoECSNet: A Local Expansion and Cross-Row Sampling Self-Attention Network for 3D Object Detection

Scheduled for presentation during the Video Session "On-Demand Video Presentations" (VP-VP), Saturday, November 22, 2025, 08:00−18:00, On-Demand Platform

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on April 2, 2026

Abstract

Autonomous driving systems integrate 3D object detection to enhance the efficiency and safety of intelligent transportation systems. LiDAR-based 3D object detection plays a crucial role in autonomous driving systems. The sparse distribution of points in 3D scenes typically causes long-range feature dependency loss, which existing high-performance 3D object detection methods typically address by redesigning backbone networks or using sparse detectors. However, previous methods exhibit significant limitations in utilizing context and balancing time and space, resulting in unsatisfactory detection outcomes. To address these issues, we propose LoECSNet, a Local Expansion and Cross-Line Sampling Self-Attention Network for 3D object detection. We first designed a local partitioning expansion(LE) block for generating expansion windows. After that, the spatial-content adaptive interaction (SAI) block transforms the BEV features within these expansion windows into adaptive interaction features that incorporate contextual information. Finally, the Cross-row Self-Attention downsampling (CSA) block captures long-range dependencies between features in space, and merges them with the previously obtained adaptive interaction features for detection. We conducted extensive experiments on the nuScenes dataset, where LoECSNet outperformed previous state-of-the-art methods, demonstrating superior performance in detecting large, long-range, and contextually relevant objects.