Visual Recognition beyond the Closed World | Doctoral Program - Information Engineering and Computer Science

Visual Recognition beyond the Closed World

Massimiliano Mancini

Code:

38016

Period:

Monday, August 28, 2023 to Friday, September 1, 2023

Year:

2023

Credits:

3.00

Duration:

Objectives

Deep learning models achieved impressive results in many computer vision tasks. However, these successes heavily rely on a simple assumption: the training set captures all the knowledge the model needs to perform well at test time. In this course, we will discuss fundamental techniques to extend visual models beyond the closed world depicted in the training set, working in the open world, i.e., when test images contain different distributions or even different semantics w.r.t. training ones.

Program

The first part of the course will discuss how to overcome shifts in the input distribution (e.g., from photos to sketches), focusing on the tasks of domain adaptation, domain generalization, and their variants. In the second part, we will discuss ways to extend pre-trained models with new semantic categories, covering algorithms in incremental learning, few-shot learning, and zero-shot learning. We will then conclude with an overview of recent trends merging the two fields and related open challenges.

Schedule:

August 28 - September 1, 2023 | M T W Th F | 11 a.m. - 1 p.m.; 2 p.m. - 4 p.m..

The lectures will be held in room Garda (Povo1, second floor).