Single-view 3D Reconstruction Based on Dual Attention Mechanism

• 1、School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044

Abstract：Reconstructing a 3D shape from a single 2D image is a crucial but challenging task, whose difficulty mainly lies in depth ambiguity. In this paper, we propose a novel cyclic framework that employs silhouette information to reconstruct a 3D shape. Specifically, we first introduce position and channel attention into the reconstructor to strengthen the expression of features to reconstruct more accurate 3D shapes. Additionally, we feed the silhouettes of generated shapes back into the reconstructor along with the input images to form a complete cycle. Thus the training can be supervised via exploiting the cycle consistency, which we call silhouette cycle consistency. The experimental results on ShapeNet datasets show that the proposed method outperforms the state-of-the-art methods without relying on 3D annotations and multi-view supervision.

Keywords： 3Dobject reconstruction Single view supervision Dual attention mechanism Silhouette