[3B3-E-2] Machine learning: image recognition and generation

[3B3-E-2-01] Design a Loss Function which Generates a Spatial configuration of Image In-betweening

〇Paulino Cristovao¹, Hidemoto Nakada^1,2, Yusuke Tanimura^1,2, Hideki Asoh² (1. University of Tsukuba, 2. National Advanced Institute of Science and Technology of Japan (AIST))

Keywords:Image Inbetween, Variational Autoencoders, Deep Learning

Instead of generating image inbetween directly from adjacent frames, we propose a method based on inbetweening in latent space. We design a simple loss function which generates a latent space that represent the spatial configu- ration of image inbetween. Contrary to the frame based methods, this model can make plausible assumption about the moving objects in the image and can capture what is not seen in the images. Our model has three networks, all based on variational autoencoder, sharing same weights. We validate this model on different synthetic datasets. We show the details of our network architecture and the evaluation results.