Improving the Performance of Multi-View 3D Reconstruction Model Through Score-Dropout that Uses the Different Viewpoints

In recent years, as the deep learning-based technology has expanded in various fields, three-dimensional (3D) data reconstruction has also become an active area of research to improve performance by applying deep learning-based methods. Major deep learning-based 3D reconstruction techniques include 3D point cloud extraction, function space applications, and multi-view stereo models. In this paper, we introduce a score dropout technique, which is the novel approach to improve the performance of multiview 3D reconstruction models. This technique compares the feature values extracted during the twodimensional (2D) encoding process from multi-view images, specifies features with relatively low impact on the 3D reconstruction, and selectively dropouts them to improve performance. The main advantage of this method is that it does not require any additional learning parameters, but can enhance performance. Experiments applying this technique to the 3D-R2N2 model show that the 3D-GRU-3 model incorporating score dropout outperforms the Res-GRU-3 model with 14.3M more training parameters at five view. These results suggest that multi-view 3D reconstruction models can be expected to improve performance by incorporating score dropout.