Self-Supervised Ground-Relative Pose Estimation

Bruce R. Muller, William A. P. Smith

Abstract: We propose a self-supervised method for relative pose estimation. Unlike existing self-supervised methods, we do not train a dense depth estimation network in conjunction with our pose network and hence avoid the complexity and ambiguity of this much harder problem. Instead, we use a very simple geometric model in which we assume the local road scene is planar. By estimating a 9D ground-relative pose, we are able to perform cross-projection between images via the ground plane using only a homography to compute a self-supervised appearance loss between overlapping images. We use a geometric matching architecture that can handle arbitrary pose pairs and use a pretrained feature extractor to compute a perceptual appearance loss. Our approach is competitive with more complex visual odometry methods that estimate dense depth maps. Code: github.com/brucemuller/homographyVO