Emergence of Object Segmentation in Perturbed Generative Models

摘要: We introduce a novel framework to build model that can learn how segment objects from collection of images without any human annotation. Our method builds on the observation location object segments be perturbed locally relative given background affecting realism scene. approach is first train generative layered The representation consists image, foreground image and mask foreground. A composite then obtained by overlaying masked onto background. trained in an adversarial fashion against discriminator, which forces produce realistic images. To force generator where layer corresponds object, we perturb output introducing random shift both Because unaware before computing its output, it must representations are for such perturbation. Finally, defining autoencoder consisting encoder, train, pre-trained as decoder, freeze. encoder maps feature vector, fed input give matching original image. outputs explicit scene, learns detect objects. demonstrate this real several categories.