Context-Based Visual Object Segmentation

Abstract

In this thesis, we aim to solve the problem of object segmentation. It has been proved that both classification and detection can provide useful contextual information to guide the segmentation process. We first proposed a detection-based method that formulates the segmentation task as pursuing the optimal latent mask inside the bounding box via sparse reconstruction. Furthermore, we proposed an approach based on detection without any additional segment annotation. Finally, besides global classification and detection, we explore the contextual cues from the unlabeled background regions that are usually ignored. The proposed approaches achieve new state-of-the-art performance on various benchmark datasets, like PASCAL VOC, Weizman Horse, Grabcut-50 and MSRC-21.