Tianyi Guan

Abstract

Weakly supervised learning can adjust model parameters by exploring image-level annotations, thus significantly reducing the cost of labeling samples. Although many efforts have been made to reduce the performance gap compared to supervised learning with expensive annotations, this article focuses on overlapped objects' adverse effects in recognizing overlapped objects via weakly supervised methods. In order to identify overlapped objects by weak supervision method, the current Bag-of-Words (BOW) framework for non-overlapped objects is modified to Bag-of-Alphabets (BOA), the words of objects corresponding to the alphabets of its parts. As a result, the un-covered factors of objects can significantly contribute to object recognition even when others may cover them. So, compared to BOW, a novel method, BOA is created to combine alphabets of parts into words of objects.

Keywords: Weakly Supervised Learning; Image-Level; Recognizing Overlapped Objects; Bag-of-Words; Bag-of-Alphabets