Leveraging Equivariant Features for Absolute Pose Regression

Pose estimation enables vision-based systems to refer to their environment, supporting activities ranging from scene navigation to object manipulation. However, end-to-end approaches, that have achieved state-of-the-art performance in many perception tasks, are still unable to compete with 3D geometry-based methods in pose estimation. Indeed, absolute pose regression has been proven to be more related to image retrieval than to 3D structure. Our assumption is that statistical features learned by classical convolutional neural networks do not carry enough geometrical information for reliably solving this task. This paper studies the use of deep equivariant features for end-to-end pose regression. We further propose a translation and rotation equivariant Convolutional Neural Network whose architecture directly induces representations of camera motions into the feature space. In the context of absolute pose regression, this geometric property allows for implicitly augmenting the training data under a whole group of image plane-preserving transformations. Therefore, directly learning equivariant features efficiently compensates for learning intermediate representations that are indirectly equivariant yet data-intensive. Extensive experimental validation demonstrates that our lightweight model outperforms existing ones on standard datasets.