Full text
| |
| Author(s): |
Caroline Mazini Rodrigues
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2020-06-04 |
| Examining board members: |
Zanoni Dias;
Moacir Antonelli Ponti;
Hélio Pedrini
|
| Advisor: | Anderson de Rezende Rocha; Zanoni Dias |
| Abstract | |
Different events, such as terrorist acts and natural catastrophes, frequently occur across the world. The availability of images on the internet can help to understand events. When dealing with images from events, filter the images is one of the major challenges. The crucial data, which could indeed represent the event, might be mixed with even more massive amounts of non-important data. However, manually selecting representative (helpful) images from a massive amount of data can be infeasible. Hence, the question becomes: How to automatically separate the representative images from the non-representative ones? In this research, we propose techniques to deal with this question considering the lack of labeled images to indicate representativeness. We cope with the representativeness image retrieval using methods of Content-Based Image Retrieval - CBIR, posteriorly improved by quality ranking metrics. Nevertheless, one of the biggest problems when retrieving images is to correctly represent these images semantically. In order to propose representations which could capture the event semantics we present two approaches. Our approaches are based on representations of components which could encode the information necessary to describe the events, such as people attending it (e.g. suspects or victims); objects that appear in the event scene (e.g., cars,gun, backpack); and the place where the event unfolded (e.g.,park, stadium, building). The first approach proposed, called Event Semantic Space, intend to describe images as a low-dimensional representation using a small quantity of known representative images. The second approach, called Event Combined Space, aims to overcome the precision results of the first one by learning a manner to combine the representative components. Results on three real-world event datasets attest the capability of our methods to represent events based on representative components combination (AU) | |
| FAPESP's process: | 18/16214-3 - Heterogeneous data analysis for event detection |
| Grantee: | Caroline Mazini Rodrigues |
| Support Opportunities: | Scholarships in Brazil - Master |