基于轴向空间注意力和中间融合表示的单图像三维重建模型

3D reconstruction model based on attention and intermediate fusion representation

摘要:

利用“编码-解码”模型实现单图像的三维重建，由于特征提取能力不足导致重建对象细节缺失，针对这一问题，本文提出基于注意力和中间融合表示的三维重建模型，旨在重建具有精细化的三维模型。利用轴向空间注意力机制学习不同方向的信息，将其嵌入至残差单元中以捕获局部结构特征；基于双流网络推测深度图和三维平均形状以设计中间融合表示模块，有效融合可见表面细节信息，更好地描绘对象的三维空间结构。实验结果表明：轴向空间注意力机制和中间融合表示模块增强了特征提取的能力，IoU和F-score比PixVox++分别提升了1.3%和0.4%，三维重建效果更优。

Abstract:

The coding-decoding model is used to realize the 3D reconstruction of a single image. Due to the lack of feature extraction ability, the details of the reconstructed object are missing. To solve this problem, this paper proposes a 3D reconstruction model based on attention and intermediate fusion representation, aiming at the reconstruction of a fine 3D model.The axial spatial attention mechanism is used to learn information from different directions and embed it into residual elements to capture local structural features. Based on the dual-flow network, the depth map and 3d average shape are deduced to design the intermediate fusion representation module, which effectively integrates the visible surface details and better describes the 3d spatial structure of the object. The experimental results show that the axial spatial attention mechanism and the intermediate fusion representation module enhance the ability of feature extraction, IoU and f-score are improved by 1.3% and 0.4% respectively compared with PixVox++, and the 3d reconstruction effect is better.