基于注意力机制的单视图三维重建

Single-view 3D reconstruction based on attentional mechanisms

利用深度学习的方法进行单视图三维重建时,网络中的外部辅助结构大幅提高了其重建效果,但也增加了网络的参数量和运算量。对此,在编码器中减少卷积层数量,并设计注意力模块。在注意力模块中,利用大核注意力为图像空间信息分配不同的权重。此外,利用向量融合保持空洞卷积过程中的空间信息关系。在重建过程中,编码器通过卷积层和注意力模块交替提取图像特征后,将编码向量直接输入到解码器,在解码器中上采样并输出重建模型,实现无需外部辅助结构的重建。在ShapeNet数据集上的对比实验表明网络在较低的模型参数量和运算量下具有更好的三维重建效果。

When using the deep learning approach for single-view 3D reconstruction,the external auxiliary structure in the network substantially improves its reconstruction effect,but also increases the number of parameters and the amount of operations in the network.In response,the number of convolutional layers is reduced in the encoder and the attention module is designed.In the attention module,different weights are assigned to the image spatial information using large kernel attention.In addition,vector fusion is used to maintain the relationship of spatial information during the convolution of voids.In the reconstruction process,after the encoder extracts the image features through the convolution layer and the attention module alternately,the encoded vector is directly input to the decoder,where the reconstruction model is up-sampled and output to achieve reconstruction without external auxiliary structures.Comparative experiments on the ShapeNet dataset show that the network has better 3D reconstruction results with lower model parametric quantities and operations.