A Multi-view Pixel-wise Voting Network for 6DoF Pose Estimation

6DoF pose estimation is an important task in the Computer Vision field for what regards robotic and automotive applications. Many recent approaches successfully perform pose estimation on monocular images, which lack depth information. In this work, the potential of extending such methods to a multi-view setting is explored, in order to recover depth information from geometrical relations between the views. In particular two different multi-view adaptations for a particular monocular pose estimator, called PVNet, are developed, by either combining monocular results on the individual views or by modifying the original method to take in input directly the set of views. The new models are evaluated on the TOD transparent object dataset and compared against the original PVNet implementation, a depth-based pose estimation called DenseFusion, and the method proposed by the authors of the dataset, called Keypose. Experimental results show that integrating multi-view information significantly increases test accuracy and that both models outperform DenseFusion, while still being slightly surpassed by Keypose.