Model-based occlusion disentanglement for image-to-image translation
This addresses image quality issues in computer vision applications such as autonomous driving or surveillance, but it is incremental as it builds on existing adversarial pipelines.
The paper tackles the problem of image-to-image translation being affected by occlusions like raindrops or dirt by proposing an unsupervised model-based learning method that disentangles scene and occlusions, and it demonstrates highly realistic translations that outperform state-of-the-art methods on multiple datasets.
Image-to-image translation is affected by entanglement phenomena, which may occur in case of target data encompassing occlusions such as raindrops, dirt, etc. Our unsupervised model-based learning disentangles scene and occlusions, while benefiting from an adversarial pipeline to regress physical parameters of the occlusion model. The experiments demonstrate our method is able to handle varying types of occlusions and generate highly realistic translations, qualitatively and quantitatively outperforming the state-of-the-art on multiple datasets.