Beyond Photometric Constraints: Epipolar-Based Self-Supervised Learning for Visual Odometry

This paper introduces an enhanced self-supervised learning (SSL) framework for Visual Odometry (VO), aiming to improve pose and depth estimation in dynamic and complex environments. The proposed approach integrates three key innovations: a unified loss function, selective photometric loss, and adaptive geometric constraints. The unified loss dynamically adjusts the balance between photometric and geometric terms based on scene dynamics, ensuring accurate depth and pose predictions. The selective photometric loss focuses only on stable texture regions, reducing errors from reflections, occlusions, and moving objects. Adaptive geometric constraints refine pose estimation by leveraging robust key point matching and epipolar geometry to minimize computational complexity while maintaining high accuracy. Our framework was evaluated extensively on the KITTI dataset, achieving an RMSE of 5.564 and an absolute relative error (Abs Rel) of 0.152, outperforming existing state-of-the-art methods across key metrics. The proposed architecture also features a streamlined encoder-decoder setup with skip connections, enabling efficient joint optimization of depth and pose while reducing network complexity. Qualitative results highlight significant improvements in depth map consistency and pose estimation, particularly in challenging scenarios with high motion and dynamic objects. This work contributes to the field by addressing key limitations in self-supervised VO, such as sensitivity to dynamic objects and computational inefficiency, while demonstrating robust performance for real-world applications. The proposed framework is scalable, computationally efficient, and suitable for real-time applications, such as autonomous driving and robotics. Future research will explore integrating additional modalities, such as optical flow, to further enhance accuracy and robustness.