Piecewise Planar 3D Reconstruction for Indoor Scenes from a Single Image Based on Atrous Convolution and Multi-Scale Features Fusion [基于空洞卷积与多尺度特征融合的室内场景单图像分段平面三维重建]

作者：Sun K.; Miao J.^*; Jiang R.; Huang S.; Zhang G.

来源：Chinese Journal of Sensors and Actuators, 2021, 34(3): 370-378.

DOI：10.3969/j.issn.1004-1699.2021.03.012

• 3D reconstruction of indoor scene
• Atrous convolution
• Conditional random field
• Deep convolutional neural network
• Multi-scale feature fusion

It is hard for pixel-level and regional-level 3D reconstruction algorithms to recover details of indoor scenes due to luminous changes and lack of texture. A piecewise planar 3D reconstruction method is proposed based on the convolution residual connection of the holes and the multi-scale feature fusion network. This model uses the shallow high-resolution detail features generated by the ResNet-101 network with the added hole convolution to reduce the loss impact of spatial information as network structure deepens on the detail reconstruction, so that this model can learn more abundant features and by coupling positioning accuracy optimized by the fully connected conditional random field(CRF)with the recognition ability of deep convolutional neural network, which keeps better boundary smoothness and details. Experimental results show that the proposed method is robust to the plane prediction of indoor scenes with complex backgrounds, the plane segmentation results are accurate, and the depth prediction accuracy can reach 92.27% on average.

• 单位
  南昌航空大学