3DReg-i-Net: Improving Deep Learning Based 3D Registration for a Robust Real-time Alignment of Small-scale Scans

Abstract

We present 3DReg-i-Net, an improved deep learning solution for pairwise registration of 3D scans, which evolves the recently proposed 3DRegNet technique by Pais et al. This is one of the very first learning based algorithm aiming at producing the co-registration of two 3D views starting solely from a set of point correspondences, which is able to perform outlier rejection and to recover the registration matrix. We evolve the original method to face the challenging scenario of quick 3D modelling at small scales through the alignment of dense 3D views acquired at video frame-rate with a handheld scanner. We improve the system tracking robustness and alignment performance with a generalized input data augmentation. Moreover, working on suboptimal aspects of the original solution, we propose different improvements that lead to a redefinition of the training loss function. When tested on the considered scenario, the proposed 3DReg-i-Net significantly outperforms the prior solution in terms of accuracy of the estimated aligning transforms.