Classification reliability of 3D shapes using neural networks in case of partial and noisy models

Deep learning is nowadays a mature technique and it is widely applied to image processing and classification. In recent years, many authors tried to extend this approach also to 3D object classification. However, most of the works in this field refers to complete models, while in many real applications the single acquisition with a vision system may only provide a partial object representation. Thus, the main goal of this work is to study the behaviour of classification neural networks when partial 3D models are considered. In particular, the analysis is focused on the classification reliability using partial point clouds, evaluating the influence of noise level and object scaling on the overall network performance. Tests are carried out both on synthetic point clouds, generated by simulation of common acquisition techniques, and on real clouds acquired by a Kinect device. This pushes towards the development of hybrid solutions, where training is made on simulated clouds and the testing takes place on real scanned objects, providing interesting suggestions for practical applications