Abstract
In this paper,we introduce a novel framework for WEakly supervised Leaming of Deep cOnvolutional neu-ral Networks(WELDON)Our method is dedicated to au-tomatically selecting relevant image regions from weak an-notations,c.g global image labels,and encompasses the following contributions.Firstly,WELDON leverages recent improvements on the Mulriple Instance Leaming paradigm,i.c.negative evidence scoring and top instance selection.Secondhy,the deep CNN is trained to optimize Average Pre-cision,and fine-tuned on the target dataset with efficient computations due to conwolutional feature sharing.A thor.ough experimental validation shows that WELDON outper-fonmns state-of -the-art results on sir different datasets.