Figure 2.

Examples of a reconstruction procedure by our image-completion network. m is a function for masking patch images and f is a function for our image-completion. In Figure 2, the moving car is not included in background images used for training our image-completion network. The image-completion network learns how to reconstruct central regions using only background images. Red and blue lines represent reconstruction of a foreground region and a background region, respectively. The background region is reconstructed from a region surrounding the masked central region. A car wheel in the foreground region is not reconstructed in the red line path. The image-completion network cannot use color and texture of the car wheel. Therefore, errors in the car wheel are emphasized.