3D Reconstruction of Indoor Scene Combining Channel and Spatial Attention

Segmentation and parameters estimation of planes are the key steps of 3D piecewise planar reconstruction from a single indoors image. Most existing works of CNN-based methods ingore the global context information and the relationship between feature channels, so it is easy to attain inaccurate segmentation of small planes and make plane details lost. resulting large errors of the parameter estimation of corresponding regions. In this paper, we propose a novel CNN model with fusion feature channel attention and spatial attention. In this model, the feature response between network channels is calibrated by using channel attention, and spatial semantic information is extracted in the encoder with combining spatial attention, which focus the network on the small planes and plane's detail. Experimental results show that the proposed method can significantly improve the accuracy of plane segmentation, and the accuracy of depth prediction reaches 93.57%, which effectively improves the performance of 3D scene reconstruction.