SCSCN: A Separated Channel-Spatial Convolution Net With Attention for Single-View Reconstruction
Jiayi Ma, Hao Zhang, Peng Yi, Zhongyuan Wang · IEEE Transactions on Industrial Electronics · 2019
Three-dimensional (3-D) object reconstruction is a challenging problem in computer vision, especially the single-view reconstruction. In this article, we propose a new 3-D reconstruction network, termed as separated channel-spatial convolution net with attention (SCSCN), which can reconstruct the 3-D shape of objects by given a two-dimensional (2-D) image from any viewpoint. Our method is a simple encoder-decoder structure, where the encoder uses separated channel-spatial convolution and separated channel-spatial attention to extract features from the input image, and the decoder recovers 3-D shapes from the features. The separated channel-spatial convolution can obtain channel information and spatial information through the channel path and spatial path separately. At the same time, in order to select a more reasonable combination of features according to the degree of contribution to the reconstruction task, channel attention and spatial attention are relevantly inserted into these two paths. As a result, the encoder can extract a strong representation of object. Quantitative experiments show that our SCSCN has a weak dependence on 3-D supervision and achieves high-quality reconstruction just under 2-D supervision, which proves the effectiveness of the encoder. In addition, we conduct the qualitative visualization experiment to confirm the rationality of the attention blocks in the feature extraction process.