An improved recurrent neural networks for 3d object reconstruction

Abstract

3D-R2N2 and other advanced 3D reconstruction neural networks have achieved impressive results, however most of them still suffer from training difficulties and detail losing, due to their weak feature extraction capability and improper loss function. This paper aims to overcome these shortcomings and defects by building a brand new model based on 3D-R2N2. The new model adopts densely connected structure as encoder, and utilizes Chamfer Distance as loss function. The aim is to enhance the learning ability of the network for complex data, meanwhile, make the focus of the whole network rest on the reconstruction of detail structures. In addition, we also made an improved decoder by building two parallel predictor branches to make better use of the feature information and boost the network’s performance on reconstruction task. Through extensive tests, the results show that our proposed model called 3D-R2N2-V2 is slightly slower than 3D-R2N2 in predicting speed, but it can be 20% to 30% faster than 3D-R2N2 in training speed and obtain 15% and 10% better voxel IoU results on both single- and multi-view reconstruction tasks, respectively. Compared with other recent state-of-the-art methods like OGN and DRC, the reconstruction effect of our approach is also competitive.