Core Viewpoint - The article discusses the development of ReCamDriving, a novel trajectory video generation method that operates without LiDAR, focusing on camera control and understanding the geometric relationship between the camera and the environment [5][34]. Group 1: Methodology - ReCamDriving aims to generate new trajectory videos using only one real driving video, addressing the high costs and complexities associated with collecting multi-trajectory video data [5][6]. - The method employs a two-stage training strategy: the first stage teaches the model basic camera movement, while the second stage introduces 3D Gaussian Splatting (3DGS) for precise geometric and perspective constraints [18][24]. - Unlike traditional methods that rely on LiDAR for geometric constraints, ReCamDriving utilizes 3DGS, which provides dense and complete scene structure information, enhancing stability in generating new trajectories [10][15]. Group 2: Experimental Results - Experimental results indicate that ReCamDriving significantly outperforms other methods in camera control accuracy and visual quality across various datasets, including Waymo and nuScenes [30][31]. - The method shows improved geometric consistency, particularly in scenarios with large lateral offsets, where other methods tend to fail due to sparse LiDAR data [30][31]. - The ReCamDriving approach maintains visual continuity in fine structural areas, such as lane markings and distant buildings, even under significant lateral shifts [31][33]. Group 3: Data Construction - The authors developed the ParaDrive dataset, which includes over 110,000 pairs of parallel trajectory videos, providing a crucial foundation for future research in trajectory generation [27]. - A novel cross-trajectory data construction strategy was introduced to address the challenge of lacking true new trajectory videos for supervision during training [22][23].