| Grant number: | 20/08770-3 |
| Support type: | Scholarships in Brazil - Master |
| Effective date (Start): | September 01, 2020 |
| Effective date (End): | August 31, 2022 |
| Field of knowledge: | Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques |
| Cooperation agreement: | Microsoft Research |
| Principal researcher: | Jurandy Gomes de Almeida Junior |
| Grantee: | Lucas Fernando Alvarenga e Silva |
| Home Institution: | Instituto de Ciência e Tecnologia (ICT). Universidade Federal de São Paulo (UNIFESP). Campus São José dos Campos. São José dos Campos , SP, Brazil |
| Company: | Universidade Estadual Paulista (UNESP). Campus de Rio Claro. Instituto de Geociências e Ciências Exatas (IGCE) |
| Associated research grant: | 17/25908-6 - Weakly supervised learning for compressed video analysis on retrieval and classification tasks for visual alert, AP.PITE |
Abstract Nowadays the use of computational vision methods on real-world problems is growing mainly by the achievement of better results on image and video recognition tasks. These results have been achieved by the evolution of machine learning methods, specifically by deep neural networks. However, these methods are usually adapted for tasks where their limits are well known, and have low flexibility for data that have been out of the problem scope. In real-world problems, we expect that classification methods should achieve equivalent results of the training step. But when used on these problems, we can see low classification performance. This happens by the fact that deep networks do not generalize well on unknown data or domain different data, which is known as category shift and domain shift problems. In front of this problem, researchers have been developing methods to avoid this gap with unknown data on two main research lines: Unsupervised domain adaptation methods and Open-set methods, also known by open-world methods. While domain adaptation methods try to align domain distributions, open-set methods should predict if some data can be recognizable by the network. The goal of this work is to study and investigate open-set methods and develop new state-of-art approaches on images and video recognition tasks. (AU) | |