Component-Aware High-Resolution 3D Object Reconstruction

Abstract

<p indent="0mm">The computational cost of the voxel-based 3D object reconstruction grows cubically with the increase of the resolution. To address this problem, the proposed component-aware 3D object reconstruction method decomposes a 3D object into several components to reconstruct the 3D object by predicting and assembling a series of components, which transforms the high-resolution 3D object reconstruction into a series of low-resolution component reconstruction. The proposed method predicts the positions of all components using a component position prediction module. Then the geometric and appearance feature of a component are fused into a joint feature with a component feature extraction module. The joint feature is utilized by a component shape reconstruction module to predict the geometry of components. Finally, all components are assembled into a high-resolution 3D object with the guidance of their positions. Experiments are performed on ShapeNet dataset using an NVIDIA 1080 Maxwell GPU with 12 GB of memory. The comparison methods include an octree-based high-resolution reconstruction method, a LSTM-based low-resolution reconstruction method and a baseline method using encoder-decoder architecture. The results of high-resolution reconstruction experiment demonstrate that the component-aware 3D reconstruction method achieves a satisfactory 3D reconstruction accuracy with a low computational cost. In the low-resolution reconstruction experiment, the proposed method also performs better and the average accuracy in 13 categories reaches 0.618.