Q-PSO: fast quaternion-based pose estimation from RGB-D images

We present a pipeline for fast object pose estimation using RGB-D images, which does not rely on image features or machine learning.We are interested in segmenting objects with large variety in appearance, from lack of texture to presence of strong textures, with a focus on the task of robotic grasping. The proposed pipeline is divided into an object segmentation part and a pose estimation part. We first find candidate object clusters using a graph-based image segmentation technique. A modified Canny edge detector is introduced for extracting robust graph edges by fusing RGB and depth information. A suitable cost function is used for building the graph, which is then partitioned using the concept of internal and external differences between graph regions. The extracted object regions are then used to initialize the 3D position of a quaternion-based Particle Swarm Optimization algorithm (Q-PSO), that fits a 3D model of the object to the depth image. The fitness function is based on depth information only and the quaternion formulation avoids singularities and the need for conversions between rotation representations. In this work we focus on the details of the GPU implementation of Q-PSO, in order to fully exploit the highly parallelizable nature of the particular implementation of the particle swarm algorithm, and discuss critic implementation details. We then test the approach on different publicly available RGB-D object datasets, and provide numeric comparisons with other state-of-the-art methods, as well as a discussion on robustness and an extension to the case of articulated objects. We show how Q-PSO offers comparable performances to current learning-based approaches,while not suffering from the problems of lack of features in objects or issues related to training, such as the need for a large training set and long training times.