Metrics for Exposing the Biases of Content-Style Disentanglement

Recent state-of-the-art semi- and un-supervised solutions for challenging computer vision tasks have used the idea of encoding image content into a spatial tensor and image appearance or "style" into a vector. These decomposed representations take advantage of equivariant properties of network design and improve performance in equivariant tasks, such as image-to-image translation. Most of these methods use the term "disentangled" for their representations and employ model design, learning objectives, and data biases to achieve good model performance. While considerable effort has been made to measure disentanglement in vector representations, currently, metrics that can characterize the degree of disentanglement between content (spatial) and style (vector) representations and the relation to task performance are lacking. In this paper, we propose metrics to measure how (un)correlated, biased, and informative the content and style representations are. In particular, we first identify key design choices and learning constraints on three popular models that employ content-style disentanglement and derive ablated versions. Then, we use our metrics to ascertain the role of each bias. Our experiments reveal a "sweet-spot" between disentanglement, task performance and latent space interpretability. The proposed metrics enable the design of better models and the selection of models that achieve the desired performance and disentanglement. Our metrics library is available at https://github.com/TsaftarisCollaboratory/CSDisentanglement_Metrics_Library.