Objectives and Contextualisation

Module Coordinator:  Dr. Ramon Baldrich Caselles

The objective of this module is to introduce the Machine learning techniques for solving computer vision problems. Machine learning deals with the automatic analisys of large scale data. Nowadays it conforms the basics of many computer vision methods, specially those related to visual pattern recognition or classification, where 'patterns' encompasses images of world objects, scenes and video sequences of human actions, to name a few.

This module presents the foundations and most important techniques for the classification of visual patterns, mainly focusing on supervised methods. Also, related topics like image descriptors and dimensionality reduction are addressed. As much as possible, all these techniques are tried and assessed on a practical project concerning scene description from pictures, toghether with the standard metrics and procedures for performance evaluation like precision-recall curves and k-fold cross-validation.

The learning outcomes are:

(a) Distinguish the main types of ML techniques for computer vision: supervised vs. unsupervised, generative vs. discriminative, original feature space vs. feature vector kernelization.

(b) Know the strong and weak points of the different methods, in part learned while solving a real pattern classification problem.

(c) Being able to use existing method implementations and build them from scratch.

The module goes in depth in two main approches to introduce ML into the image classification problem. Using: a) handcrafted image description, b) data driven image description.  On the first case the Bag of Words is used, on the second one, the Deep Learning approach. The DL content is developed extensively providing both, thoretical basis of the different parts of modern Neural Networs acrhitecutres, and best practices to apply it on real applications.