SpyroPose: SE(3) Pyramids for Object Pose Distribution Estimation
Rasmus Laurvig Haugaard, Frederik Hagelskjær, Thorbjørn Mosekjær Iversen
Abstract
Object pose estimation is an essential computer vision problem in many robot systems. It is usually approached by estimating a single pose with an associated score, however, a score conveys only little information about uncertainty, making it difficult for downstream manipulation tasks to assess risk. In contrast to pose scores, pose distributions could be used in probabilistic frameworks, allowing downstream tasks to make more informed decisions and ultimately increase system reliability. Pose distributions can have arbitrary complexity which motivates unparameterized distributions, however, until now they have been limited to rotation estimation on SO(3) due to the difficulty in training on and normalizing over SE(3). We propose a novel method, SpyroPose, for pose distribution estimation using an SE(3) pyramid: A hierarchical grid with increasing resolution at deeper levels. The pyramid enables efficient training through importance sampling and real time inference by sparse evaluation. SpyroPose is state-of-the-art on SO(3) distribution estimation, and to the best of our knowledge, we provide the first quantitative results on SE(3) distribution estimation. Pose distributions also open new opportunities for sensor-fusion, and we show a simple multi-view extension of SpyroPose. Project page at spyropose.github.io