Bilkent University
Department of Computer Engineering


Learning to Generate Multi-View Representation of Objects for Multi-View Search


Arda Ünal
MS Student
Computer Engineering Department
Bilkent University

Multiple view images, both in the queries and in the database, signi?cantly improve the retrieval precision. In this work, we generate new views of the objects from the ones we have and investigate its impacts on the retrieval precision. To generate the new views we take advantage of generative models specifically Variational Autoencoders (VAEs) as they have emerged as one of the most popular approaches to unsupervised learning of complicated distributions. VAEs have already shown promise in generating many kinds of complicated data, including handwritten digits and faces.


DATE: 24 October, 2016, Monday @ 16:40