From Objects to Actions (presentation PDF): Computer vision research has been making steady progress for the last 20 years or so. In the last few years it has made much faster advances thanks to improved hardware, an explosion in the amount of visual data available and the application of deep neural networks. An area that is still a challenge is the automatic description of human actions from video, otherwise known as human action recognition or human action detection. This talk will present some ideas on how to address this problem.

Similarity search for weakly supervised Machine Learning: The growth of the datasets terms of size and number of classes blurs the lines between the classification and instance search tasks. K-nn classification has several drawbacks that prevent it from being competitive in a classical supervised setting. However, it has advantages in a "not really" supervised setting, ie. where there are tags rather than classes, or multi-media annotations (freeform text, face ids, GPS coordinates, URLs, etc.) It also generalizes smoothly to on-line training and situations where there is no clear separation between a train and a test set. We will present a few computer vision works where similarity search is applied to unsupervised and weakly supervised learning in computer vision.

Face recognition: from traditional to learning-based methods: Starting in the seventies, face recognition has become one of the most researched topics in computer vision and biometrics. Traditional methods and machine learning techniques have over recent years been superseded by both deep learning approach based on convolutional neural networks (CNNs), and sparse dictionary learning modeled within the sparse representation paradigm. The deep neural networks methods rely heavily on the size and representative of training set and suffer or even fail with few samples per person. Instead, learning a sparse dictionary from small training sets (or at the very limit one sample per subject) turns out to be a robust method guaranteeing appreciable performances also against poor availability of information per subject. The purpose of this talk is to describe how effectively learning techniques have improved the global approach to the identity recognition problem, addressing relevant issues in this area such as data collection, open challenges, and several promising directions for future research.