04 Jun 2020
14:00 Master's Defense Fully distance
Theme
Image Representativity Analysis for Event Description
Student
Caroline Mazini Rodrigues
Advisor / Teacher
Zanoni Days
Brief summary
A variety of events, such as terrorist acts and natural disasters, occur frequently across the world. The availability of images on the internet can help to understand these events. When dealing with event images, filtering is one of the main challenges. The crucial data, which could actually represent the event, may be mixed with even greater amounts of unimportant data. However, manual selection of representative (useful) images from a large amount of data may be impracticable. So, a question arises: How to automatically separate representative and non-representative images? We proposed techniques to deal with this issue, considering the lack of images labeled to indicate representativeness. We deal with image retrieval by representation using content-based image retrieval methods (CBIR), which were subsequently enhanced by metrics for ranking quality assessment. However, one of the biggest problems encountered when retrieving images is to represent them correctly semantically. To propose representations capable of capturing the semantics of events, we present two approaches. Our approaches are based on representations of components that can encode the information necessary to describe the events, such as people who were part of the event (for example, suspects or victims); objects that appear on the scene (for example, cars or weapons); and the location where the event took place (for example, parks, stadiums or buildings). The first proposed approach, called the Semantic Space of Events, aims to describe images as a low dimensional representation using a small amount of known representative images. The second approach aims to improve the precision results of the first approach by training a combination of the representative components. The results in three real-world event data sets attest to the ability of our methods to represent events based on the combination of representative components.
Examination Board
Headlines:
| Zanoni Dias | IC / UNICAMP |
| Hélio Pedrini | IC / UNICAMP |
| Moacir Antonelli Ponti | ICMC / USP |
Substitutes:
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| Roberto Hirata Junior | IME / USP |