Title:	Deep Visual Learning
DNr:	SNIC 2015/1-464
Project Type:	SNIC Medium Compute
Principal Investigator:	Michael Felsberg <michael.felsberg@liu.se>
Affiliation:	Linköpings universitet
Duration:	2015-12-28 – 2017-01-01
Classification:	10207
Homepage:	http://www.cvl.isy.liu.se/
Keywords:

Abstract

Features learned using Convolutional Neural Networks (CNNs) have shown significant performance gains in many computer vision applications. These features (also sometimes called deep features) are extracted from hidden convolutional layers of CNNs. They are generic and result from training on a large amount of training data (e.g. ImageNet). Due to the utilization of large training data (in the range of 14 million images), these networks require significant amount of computational power (both in terms of CPU and RAM), thereby prohibiting their training on a standard machine. This project will investigate new learning algorithms for "deep architectures". The newly developed deep representations will be applied to several computer vision applications such as image classification, visual tracking, point-set registration and action recognition. The training of the proposed deep architectures will require a large RAM (minimum 128-256 GB) with GPU resources.