A Shared Pose Regression Network for Pose Estimation of Objects from RGB Images

In this paper we propose a shared regression network to jointly estimate the pose of multiple objects, replacing multiple object-specific solutions. We demonstrate that this shared network can outperform other similar approaches that rely on multiple object-specific models by evaluating it on the TLESS dataset using the VSD (Visible Surface Discrepancy). Our approach offers a less complex solution, with fewer parameters, lower memory consumption and less training required. Furthermore, it inherently handles symmetric objects by using a depthbased loss during training and can predict in real-time. Finally, we show how our proposed pipeline can be used for fine-tuning a feature extractor jointly on all objects while training the shared pose regression network. This fine-tuning process improves the pose estimation performance.