Spatial Separable Multi-scale Feature Net: A 3D Reconstruction Method with More Features but Fewer Parameters

• Published in: 2022 International Joint Conference on Neural Networks (IJCNN) pp. 1 - 8
• Main Authors: Li, Xuanang, Liang, Zhengyou, Sun, Yu, Yao, Qiang
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18.07.2022
• Subjects: 3D Reconstruction; Feature extraction; Implicit Function; Memory management; Monte Carlo methods; Neural networks; Shape; Spatial Separable Multi-scale Feature; Surface reconstruction; Three-dimensional displays
• ISSN: 2161-4407
• DOI: 10.1109/IJCNN55064.2022.9892617

In recent years, 3D reconstruction methods based on implicit functions have shown advantages, but there are still two major problems: one is that the networks of these methods have high complexity with enormous parameters and huge memory consumption; another is that some methods' reconstructed surfaces are overly smooth and lack details, and some methods suffer from poor robustness after improving the surface details. To address these problems, we propose Spatial Separable Multi-scale Feature Net (SSMF-Net), which separately extracts the features of X-Y plane domain and Z depth domain on multiple feature scales, and generates a spatial separable multi-scale feature vector with sufficient local and global features. SSMF-Net obtains a powerful implicit function that is more faithful to the ground truth surfaces, and greatly reduces the network parameters and memory consumption at the same time. Experiments performed on ShapeNet and BUFF datasets show that our SSMF-Net outperforms PSGN, DMC, OccNet, and IF-Net. SSMF-Net preserves more local features while ensuring the global shape integrity, and reduces network parameters and memory consumption by 29%~84%. Moreover, SSMF-Net has stronger robustness and generalization.