A Multi-task Learning Convolutional Neural Network for Object Pose Estimation^⋆

摘要

Estimating 6D poses of objects from RGB images is very crucial for robots to interact with the surrounding environment and to cooperate with humans. It is a challenging problem due to the various shapes of objects, the occlusions among objects, as well as the complexity of the scene. In this study, we present a new multi-task convolutional neural network for 6D object pose estimation, which also learns the object region in the image to further improve the estimation result. A weighted loss computed from the training samples is adopted to guarantee that the components of the estimated pose are equally accurate. Additionally, we sythesize new virtual data based on the existing training data so as to overcome the possible adverse situation caused by insufficiency of training data. Experiments on the YCB-video dataset are undertaken to validate the effectiveness of the proposed method.