SFEF: Transformer-based LiDAR-Camera Object Detection for Autonomous Driving

Precise 3D object detection is crucial for self-driving cars, but it is tricky. Deep convolutional networks have shown promise in combining Light Detection and Ranging (LiDAR) and camera data for this task. However, existing methods often use simple network designs and fixed 3D bounding boxes, which don't fully capture the relationships and variations in the size of 3D objects. In response, we introduce a novel fusion algorithm named Scalable Feature Extraction Scalable Fusion (SFEF) method, which considers both long-range dependencies in the features and also different object sizes by considering two scalable units termed intra-domain feature extraction and inter-domain fusion units. The results show the superiority of the proposed method in detecting several objects and achieves overall comparable results in the nuScenes leaderboard.