Exploiting textual queries for dynamically visual disambiguation

Due to the high cost of manual annotation, learning directly from web has attracted broad attention. One issue that limits performance current webly supervised models is problem visual polysemy. In this work, we present a novel framework resolves polysemy by dynamically matching candidate text queries with retrieved images. Specifically, our proposed includes three major steps: first discover and then select according keyword-based image search results, employ saliency-guided deep multi-instance (MIL) network remove outliers learn classification for disambiguation. Compared existing methods, approach can figure out right senses, adapt dynamic changes in outliers, jointly models. Extensive experiments ablation studies on CMU-Poly-30 MIT-ISD datasets demonstrate effectiveness approach.