Visual attribute disentanglement using self-supervision

In this work, a self- supervised approach that can identify the semantic axes in the latent spaces of pre-trained adversarial generative networks is presented. Thus, a novel Direction-Disentangler model is proposed. This convolutional model employs the ResNet18 feature extractor architecture in a twin network fashion. In a self-supervised manner, a dataset for training the Direction-Disentangler model is created with the images generated by a generator network. This data is used to train the Direction-Disentangler model to identify the direction in the generator's latent space that encodes the discriminative features between two reference images. New datasets were obtained by using pre-trained InfoGAN and DCGAN models for the MNIST and FreyFaces dataset. A heuristic distance metric was used to generate consistent data samples for the multimodal MNIST data. For this purpose, a pre-trained variational autoencoder and classifier were employed to create embedding vectors. The experimental results using these datasets show that the identified direction can be applied to different reference images and is effective for editing their relevant semantic attributes.