Learning-Deep-Learning

Volumetric and Multi-View CNNs for Object Classification on 3D Data

Mar 2019

tl;dr: Improvement of volumetric CNNs (3d shapenets) closes its gap with multi-view CNNs (MVCNN).

Overall impression

The paper starts with one hypothesis: the performance gap of volumetric and multi-view CNN is due to the resolution difference. However experiment shows this only explains part of the gap. The paper then takes on two directions: improve the volumetric CNN architecture, and exploit the resolution in MVCNN. This paper already shows the concise and straightforward style of Charles Qi’s style later shown in pointnet.

Key ideas

Technical details

Notes