Improved Disentanglement through Learned Aggregation of Convolutional Feature Maps

Abstract

We present and discuss a simple image preprocessing method for learning disentangled latent factors. In particular, we utilize the implicit inductive bias contained in features from networks pretrained on the ImageNet database. We enhance this bias by explicitly fine-tuning such pretrained networks on tasks useful for the NeurIPS2019 disentanglement challenge, such as angle and position estimation or color classification. Furthermore, we train a VAE on regionally aggregate feature maps, and discuss its disentanglement performance using metrics proposed in recent literature.

Tasks

Reproductions