ITSC 2025 Paper Abstract

Paper WE-EA-T4.1

Liu, Mingyu (Technical University of Munich), Hou, Yutong (Technical University of Munich), Yurtsever, Ekim (The Ohio State University), Xingcheng, Zhou (Technical University of Munich), Zhang, Jiajie (Technical University of Munich), Zhang, Haolin (Xi'an Jiaotong University), Knoll, Alois (Technische Universität München)

Surface3D: A Surface-Aware Framework for Refining 3D Object Detection

Scheduled for presentation during the Regular Session "S04b-Intelligent Perception and Detection Technologies for Connected Mobility" (WE-EA-T4), Wednesday, November 19, 2025, 13:30−13:50, Surfers Paradise 1

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 October 19, 2025

Keywords Real-time Object Detection and Tracking for Dynamic Traffic Environments, Deep Learning for Scene Understanding and Semantic Segmentation in Autonomous Vehicles

Abstract

3D object detection aims to identify objects and localize them by predicting bounding boxes from LiDAR point clouds. However, the inherent sparsity of LiDAR data, partial occlusions, and challenges posed by long-tail viewing angles often limit their performance. To address these limitations, we propose Surface3D, a surface-aware refinement approach that integrates learned surface features derived from part segmentation to enhance bounding box predictions. Existing datasets primarily focus on indoor scene segmentation and are not suitable for our task. To address this, we develop a semi-automated pipeline to generate training data for part segmentation. This module extracts object surface features from the initial detection outputs. A refinement module then leverages these features to improve the accuracy of bounding box sizes and orientations. Extensive experiments on the KITTI dataset using five widely adopted 3D object detectors demonstrate that incorporating surface features substantially improves detection performance. Our approach improves 3D object detection performance across five detectors on KITTI. Specifically, it achieves up to 1.75% and 1.25% gains over PointPillars and VoxelRCNN, respectively, on the hard car class under 3D AP R11 metrics.