Rapid Abstraction of Spacecraft 3D Structure from Single 2D Image
Tae Ha Park, Simone D’Amico
Abstract
This paper presents a Convolutional Neural Network (CNN) to simultaneously abstract the 3D structure of the target space resident object and estimate its pose from a single 2D image. Specifically, the CNN predicts from a single image of the target a unit-size assembly of superquadric primitives which can individually describe a wide range of simple 3D shapes (e.g., cuboid, ellipsoid) using only a few parameters. The proposed training pipeline employs various types of supervision in both 2D and 3D spaces to fit an assembly of superquadrics to man-made satellite structures. In order to avoid numerical instability encountered when evaluating superquadrics, this work proposes a novel, numerically stable algorithm based on dual superquadrics to evaluate a point on the surface of and inside a superquadric for all shape parameters. Furthermore, in order to train the CNN, this work also introduces a novel dataset comprising 64 different satellite models and 1,000 images, binary masks and pose labels for each model. The experimental studies reveal that the proposed CNN can be trained to reconstruct accurate superquadric assemblies when tested on unseen images of known models and capture high-level structures of the unknown models most of the time despite having been trained on an extremely small dataset.