Abstract

Weakly Supervised Object Localization (WSOL) task attracts more and more attention in recent years, which aims to locate the object by using incomplete labels. Considering the cost of annotation, especially ground-truth bounding box label and training speed of detection task, it is very necessary to improve the performance of WSOL that only requires image-level labels. Most current methods tend to utilize Class Activation Map (CAM) that can only highlight the most discriminative parts rather than the entire target. The common method to address this kind of limitation is to hide the most obvious regions and let the model learn other parts of the target. The main work of this thesis is to eliminate the limitations of current WSOL work and improve the performance of localization. In chapter 3, we design an attention-based selection strategy to dynamically hide the feature maps. In chapter 4, a new hiding method is proposed to further improve the localization performance. In chapter 5, we propose three method to eliminate the issues on CAM level. Our methods are evaluated on CUB-200-2011 and ILSVRC 2016 datasets. Experiments demonstrate that the proposed methods work very well and significantly improve the localization performance.