Self Paced Deep Learning for Weakly Supervised Object Detection

In a weakly-supervised scenario, object detectors need to be trained using image-level annotation only. Since bounding-box-level ground truth is not available, mostof the solutions proposed so far are based on an iterative approach in which theclassifier, obtained in the previous iteration, is used to predict the objects’ positionswhich are used for training in the current iteration. However, the errors in thesepredictions can make the process drift. In this paper we propose a self-paced learning protocol to alleviate this problem.The main idea is to iteratively select a subset of samples that are most likely correct,which are used for training. While similar strategies have been recently adoptedfor SVMs and other classifiers, as far as we know, we are the first showing that aself-paced approach can be used with deep-net-based classifiers. We show results on Pascal VOC and ImageNet, outperforming the previous stateof the art on both datasets and specifically obtaining more than 100% relativeimprovement on ImageNet.