Superpixels and Attention for High-quality Object Proposals in Complex Environments

Abstract

The class-agnostic discovery of objects in images, known as object proposal generation, is a fundamental task in computer vision. In this thesis, we mainly address two major challenges in object proposal generation. First, we introduce a system that improves the challenging discovery of small objects. Second, we propose a refinement method that allows a more accurate segmentation of the discovered objects. Overall, this leads to high-quality object proposals for objects of all sizes.