09 out 2020
15:00 Master's Defense Fully distance
Theme
Multiple Image-domain Saliency Estimation using a Flexible Framework
Student
Leonardo de Mélo João
Advisor / Teacher
Alexandre Xavier Falcao
Brief summary
Protruding object detection estimates the objects that stand out most in an image. Unsupervised salience estimators use a predetermined set of assumptions about how humans perceive salience to identify salient object discriminating characteristics. Since these methods fix these predetermined assumptions as an integral part of your model, these methods cannot be easily extended to specific scenarios or other image domains. We then propose an iterative framework for salience estimation based on superpixels, called ITSELF (Iterative Saliency Estimation fLexible Framework). Our framework allows the user to add multiple overhang assumptions to better represent their model. Thanks to advances in segmentation algorithms by superpixels, boss maps can be used to improve the design of superpixels. Combining superpixel algorithms based on boss information with boss estimation algorithms based on super pixels, we propose a cycle for iterative self improvement of boss maps. We compared the ITSELF with two other state-of-the-art salience estimators in five metrics and six data sets, four of which are composed of natural images, and two are composed of biomedical images. The experiments show that our approach is more robust when compared to other methods, presenting competitive results in natural images and surpassing them in biomedical images.
Examination Board
Headlines:
| Alexandre Xavier Falcão | IC / UNICAMP |
| Silvio Jamil Ferzoli Guimarães | ICEI / PUC Minas |
| Hélio Pedrini | IC / UNICAMP |
Substitutes:
| Gerberth Adín Ramírez Rivera | IC / UNICAMP |
| Paulo André Vechiatto de Miranda | IME / USP |