Object discovery via layer disposal

Abstract

A key limitation of semantic image segmentation approaches is that they require large amounts of densely labeled training data. In this thesis, we introduce a method to learn to segment images with unlabeled data. The intuition behind the approach is that removing objects from images will yield natural images, however removing random patches will yield unnatural images. We capitalize on this signal to develop an auto-encoder that decomposes an image into layers, and when all layers are combined, it reconstructs the input image. However, when a layer is removed, the model learns to produce a different image that still looks natural to an adversary, which is possible by removing objects. Experiments and visualizations suggest that this model automatically learns to segment objects in images better than baselines. Some parts of this thesis represent joint work with Dr. Carl Vondrick and Professor Antonio Torralba.