Unsupervised Object Localization using Generative Adversarial Networks
share
This paper introduces a novel end-to-end deep neural network model for unsupervised object localization for the first time. We adopt Generative Adversarial Networks (GANs) to object localization framework, where GANs are powerful tools that implicitly learn the unknown data distribution in an unsupervised manner. It is because we observe that GAN discriminator is highly influenced by pixels where objects appear. Motivated by this observation, we apply an existing technique that visualizes important pixels for classification to the discriminator. Based on extensive evaluations and experimental studies, the proposed approach achieves meaningful accuracy for unsupervised object localization using public benchmark datasets, even comparable to state-of-the-art weakly-supervised approach.
READ FULL TEXT