Weakly Supervised Object Localization with Noisy-Label Learning

Yuming Fan, Shikui Wei, Chuangchuang Tan, Xiaotong Chen, Yao Zhao

Abstract: A novel perspective for Weakly Supervised object localization is proposed in this paper. Most recent pseudo-label-based methods only consider how to get better pseudo-labels and do not consider how to apply these imperfect labels properly. We propose the Noisy-Label Learning on Weakly Supervised object localization (NL-WSOL) to improve localization performance by cleaning defective labels. First, we generate labels which more focused categories for images in the label generation stage. Then, we judge the quality of pseudo labels and enhance the labels with poor quality. Moreover, we introduce a composite loss function to guide the network training in the pseudo-label-based training phase. Our method achieves 97.39% localization performance on the CUB-200–2011 test set.