Optimizing Depth and Ego-Motion Estimation with Swin Transformer: A Comprehensive Study

Accurately perceiving depth and motion within single images is crucial in computer vision, especially for applications like autonomous vehicles and virtual environments. While convolutional neural networks (CNNs) have shown effectiveness, their limitation in capturing wider contexts restricts current network topologies to local processing, hindering the potential of self-supervised methods. This study introduces a novel method utilizing the Swin Transformer to extract precise information from images of varying sizes, employing advanced techniques such as window and shifted-window-based attention mechanisms. Coupled with an efficient decoder network, this enhances depth estimation. Extensively tested on datasets like KITTI, this method consistently outperforms traditional approaches in predicting depth and motion accurately. Challenges remain, notably in discerning obscured items and navigating complex environments, compounded by limited dataset evaluations. Nevertheless, this approach shows promise for practical deployment, given its adaptability to diverse scenarios, underscoring the significant potential of leveraging the Swin Transformer for improved depth and motion estimation in computer vision systems.