Learning Local Image Descriptors with Autoencoders

Abstract

In this paper, we propose an efficient method for learning local image descriptors with convolutional autoencoders. We design an autoencoder architecture that yields computationally efficient extraction of patch descriptors through an intermediate image representation. The proposed approach yields significant savings in memory and processing time compared to a reference autoencoder-based patch descriptor. The results demonstrate improved robustness to noise and missing data.