The 8-Point Algorithm as an Inductive Bias for Relative Pose Prediction by ViTs

We present a simple baseline for directly estimating the relative pose (rotation and translation, including scale) between two images. Deep methods have recently shown strong progress but often require complex or multi-stage architectures. We show that a handful of modifications can be applied to a Vision Transformer (ViT) to bring its computations close to the Eight-Point Algorithm. This inductive bias enables a simple method to be competitive in multiple settings, often substantially improving over the state of the art with strong performance gains in limited data regimes.

Automated summary

We propose a simple baseline for estimating the relative pose between two images, which can directly compute the rotation, translation, and scale. By making a few modifications to the Vision Transformer (ViT), we are able to achieve results close to the Eight-Point Algorithm. This approach provides a straightforward method that is highly competitive in various scenarios, especially in cases with limited data.