Get the PhD thesis manuscript.

Get the slides of the presentation.

Watch the video of the presentation.

The talk, titled Introduction to deep learning. The techniques enabling the current Artificial Intelligence revolution, will take place at 11:45 at the Adela Moyua room of the faculty, in the campus of Leioa.

The paper, shared work between Tecnalia and the Computer Vision Center/Universitat Autònoma de Barcelona, makes use of synthetic blurring to show how we can use self-supervised and weakly supervised learning techniques to train a Convolutional Neural Net on the task of segmenting the defocus or motion-blurred areas of an image without having access to images annotated under real blur.

More info and full text available here.

This is shared work between Tecnalia’s Computer Vision Group and the Research group on Bioengineering and Resuscitation (Biores) from the University of the Basque Country (UPV/EHU). The paper shows how, by applying a neural network comprising convolutional and LSTM modules, we can successfully detect lethal ventricular arrhythmia on Out-of-Hospital Cardiac Arrest (OHCA) patients.

More info and full text available here.

The paper, presented in collaboration with BASF, extends our previous work, and describes how we migrated our classical computer vision workflow-based disease classification engine to a Deep Convolutional Neural Networks-based solution, exhaustively tested in real field-conditions.

More info and full text available here soon.

Our paper On the Duality Between Retinex and Image Dehazing has been accepted to CVPR 2018, to be held next June in Salt Lake City. The camera ready version of the paper is now available on arxiv through this link.

In this work, we prove that image enhancement algorithms based on the retinex color vision model and dehazing algorithms based on Kochsmieder’s model are related at a fundamental, modelling level. Extensive experiments show how we can obtain state of the art image dehazing results using retinex implementations over images with inverted intensities, and inverting back the result.

Tomorrow I’ll be giving an introductory lecture to Deep Learning techniques at the School of Engineering of Bilbao, University of the Basque Country. Back to where it all started!

The slides of our presentation at the Physics-Based vision meets Deep Learning (PBDL) workshop at ICCV 2017, corresponding to our paper Adversarial Networks for Spatial Context-Aware Spectral Image Reconstruction from RGB, are now avaiable here.

In addition, we also provide the ICVL dataset train-test splits used during the experimental evaluation. We have updated the arxiv version of the paper, which now shows this information in the supplementary material section, and updated the quantitative results in Table 1 for other’s methods.

Our paper Adversarial Networks for Spatial Context-Aware Spectral Image Reconstruction from RGB, has been accepted for oral presentation at the ICCV 2017 Workshop on Physics-Based Vision meets Deep Learning, to be held next October 23rd in Venice, Italy.

The paper shows how we can use the framework provided by the Generative Adversarial Networks (GANs) in a conditional setting in order to accurately reconstruct a hyperspectral image composed by 31 channels, taking only an RGB image as input.

Full program and schedule for the workshop can be found here, and a detailed pdf version with the program for all the workshops is available here