Contrastive Late Fusion for 3D Object Detection

Abstract

In the field of autonomous driving, accurate and efficient 3D object detection is crucial for ensuring safe and reliable operation. This paper focuses on the fusion of camera and LiDAR data in a late-fusion manner for 3D object detection. The proposed approach incorporates contrastive learning to enhance feature consistency between camera and LiDAR candidates, which is named as Contrastive Camera-LiDAR Object Candidates (C-CLOCs) fusion network, facilitating better fusion results. We delve into the label assignment aspect in late fusion methods and introduce a novel label assignment strategy to filter out irrelevant information. Additionally, a Multi-modality Ground-truth Sampling (MGS) method is introduced, which leverages the inclusion of point cloud information from LiDAR and corresponding images in training samples, resulting in improved performance. Experimental results demonstrate the effectiveness of the proposed method in achieving accurate 3D object detection in autonomous driving scenarios.