β-VARIATIONAL AUTOENCODERS FOR LEARNING INVERTIBLE LOCAL IMAGE DESCRIPTORS

In this paper, we propose an e ffi cient method for learning a local image descriptor and its inversion function using a modified version of a variational autoencoder (VAE) - a β -VAE. We examine di ff erent values of β in the loss function of the β -VAE to find the an optimal balance between incentivising the similarities between input patches to be preserved in latent space, and ensuring good reconstruction of the patches from their encodings in latent space. Our proposed descriptor demonstrates patch retrieval comparable to the reference autoencoder-based local image descriptor, and also shows improved reconstruction of patches from their encodings.