Supervisors: Prof. Dr. Ender Konukoglu, Prof. Dr. Minsu Cho, Prof. Dr. Bohyung Han
Object localization in images have been an interesting topic of computer vision. Many object localization methods rely on a significant amount of annotated images. However, unsupervised approaches is an open problem. In this work, we adopt an unsupervised approach, where we combine generative adversarial networks (GANs) and spatial transformer (ST). GANs, were introduced as a successful way for training generative models. This model is practiced for image completion, where occluded parts of the input image are in-painted. ST is introduced for training robust to achieve correct classification. We propose a model to detect object during image completion. The generator learns to complete the occluded image and ST learns to occlude the region hard to complete for the generator. This way we detect the challenging region in image which happens to be the object. Our method detects the object and creates bounding box under some constraints successfully.