Latent Optimization for Non-adversarial Representation Disentanglement

Disentanglement between pose and content is a key task for artificial intelligence and has attracted much research interest. Current methods for disentanglement include adversarial training and introducing cycle constraints. In this work, we present a novel disentanglement method which does not use adversarial training, achieving state-of-the-art performance. Our method uses latent optimization of an architecture borrowed from style-transfer, to enforce separation of pose and content. We overcome the test generalization issues of latent optimization, by a novel two-stage approach. In extensive experiments, our method is shown to achieve better disentanglement performance than both adversarial and non-adversarial methods that use the same level of supervision.