Ray Visual Odometry

Abstract

Learning-based Visual Odometry (VO) has seen significant advancements over the past decades. However, all the existing methods rely on the six degrees of freedom (6-DoF) representation for pose prediction, which is sparse and less conducive for neural network learning. In this work, we introduce a novel dense and distributed representation by modeling VO as ray bundles, referred to as RayVO. This richly parameterized representation is tightly coupled with corresponding spatial features, making it highly effective for neural learning. Additionally, the ray-based approach enables simultaneous prediction of both intrinsic and extrinsic parameters. To prove its effectiveness against the traditional 6-DoF representation, we propose three specialized loss functions for ray’s training: a ray-based loss, a 6-DoF-based loss and a hybrid loss. We extensively evaluate RayVO on both indoor and outdoor benchmark datasets and show that it outperforms the state-of-the-art VO methods.