6D Pose Estimation for Industrial Applications

Object pose estimation is important for systems and robotsto interact with the environment where the main challenge of this taskis the complexity of the scene caused by occlusions and clutters. A keychallenge is performing pose estimation leveraging on both RGB anddepth information: prior works either extract information from the RGBimage and depth separately or use costly post-processing steps, limitingtheir performances in highly cluttered scenes and real-time applications.Traditionally, the pose estimation problem is tackled by matching featurepoints between 3D models and images. However, these methods requirerich textured models. In recent years, the raising of deep learning hasoffered an increasing number of methods based on neural networks, suchas DSAC++, PoseCNN, DenseFusion and SingleShotPose. In this work,we present a comparison between two recent algorithms, DSAC++ andDenseFusion, focusing on computational cost, performance and applica-bility in the industry