Self Paced Deep Learning for Weakly Supervised Object Detection.

In a weakly-supervised scenario object detectors need to be trained usingimage-level annotation alone. Since bounding-box-level ground truth is notavailable, most of the solutions proposed so far are based on an iterative,Multiple Instance Learning framework in which the current classifier is used toselect the highest-confidence boxes in each image, which are treated aspseudo-ground truth in the next training iteration. However, the errors of animmature classifier can make the process drift, usually introducing many offalse positives in the training dataset. To alleviate this problem, we proposein this paper a training protocol based on the self-paced learning paradigm.The main idea is to iteratively select a subset of images and boxes that arethe most reliable, and use them for training. While in the past few yearssimilar strategies have been adopted for SVMs and other classifiers, we are thefirst showing that a self-paced approach can be used with deep-network-basedclassifiers in an end-to-end training pipeline. The method we propose is builton the fully-supervised Fast-RCNN architecture and can be applied to similararchitectures which represent the input image as a bag of boxes. We showstate-of-the-art results on Pascal VOC 2007, Pascal VOC 2010 and ILSVRC 2013.On ILSVRC 2013 our results based on a low-capacity AlexNet network outperformeven those weakly-supervised approaches which are based on much higher-capacitynetworks.