Method for creating large datasets for deep learning to improve image depth accuracy

Murayama, Masahiro*; Harazono, Yuki*; Ishii, Hirotake*; Shimoda, Hiroshi*; Taruta, Yasuyoshi   

We compared the depth enhancement process and the post-processing of the proposed and existing dataset creation methods. Results show that our depth enhancement process can create a higher quality dataset than that created using the existing method. A network trained on a dataset with our post processing completed the missing area of depth image more correctly and improves accuracy near edges better than the existing method. We also evaluated some post-processing steps for depth enhancement of the trained network.