VSTFusion-VO: Monocular Visual Odometry with Video Swin Transformer Multimodal Fusion

This paper presents a learning-based monocular visual odometry (VO) framework that leverages a Video Swin Transformer for hierarchical 3D spatiotemporal modeling. Beyond incorporating pseudo-depth, our method employs early multimodal fusion and 3D patch embedding to jointly encode RGB and geometric information before transformer-based processing, enabling effective spatiotemporal representation learning without relying on ground-truth depth. Trained end-to-end to predict 6-DoF poses, the model captures both local motion patterns and long-range dependencies. Experiments on the KITTI Odometry dataset demonstrate superior performance compared to prior learning-based VO methods in both translation and rotation accuracy. The code is available at: https://github.com/tongyu0924/VSTFusion-VO