A Recurrent and Meta-learned Model of Weakly Supervised Object Localization

The object localization and detection has improved greatly over the past decade, thanks to developments in deep learning based representations and localization models. However, a major bottleneck remains at the reliance on fully-supervised datasets, which can be difficult to gather in many real-world scenarios. In this work, we focus on the problem of weakly-supervised localization, where the goal is to localize instances of objects based on simple image-level class annotations. In particular, instead of engineering a specific weakly-supervised localization model, we aim to meta-learn a recurrent neural network based model that aims to take series of training images of a novel class, and progressively discover the foreground pattern over them. We experimentally explore the model over scenes composed of MNIST digits and noisy patches as distractors.