Attilio Fiandrotti
University of Turin (Italy)
Around and beyond the variational model
Language: English
Date: October 7th – 13h30
Abstract
Learned image compression is enjoying tremendous success, boasting compression efficiency equal or better than standardized codecs. At the basis of such success, is the the convolutional variational autoencoder, more recently also transformer-based, empowered by sophisticated entropy models.
Still, standardized codecs hold an edge in terms of ability to control the rate-distortion tradeoff and the complexity of the decoding process: in the first part of this talk, we will discuss different approaches to tackle these two key issue and close the gap with standardized codecs.
In the second part of this talk, we will explore recent trends in learned image compression that depart from the variational model, namely the adversarial and more recently diffusion models. While codecs based on such models promise to achieve ultra low bitrates that other models cannot attain, they come with some issues that we will discuss in this part of the talk.
Biography
Attilio Fiandrotti received his M.Sc. and his PhD degree in Computer Science from Politecnico di Torino in 2005 and 2010 respectively. His research interests include video coding and video streaming over wireless networks, peer-to-peer based distribution of video contents with network coding, image analysis with traditional and deep machine learning techniques.