BRULÈ: Barycenter-Regularized Unsupervised Landmark Extraction

Unsupervised retrieval of image features is vital for many computer vision tasks where the annotation missing or scarce. In this work, we propose a new unsupervised approach to detect landmarks in images, validating it on popular task human face key-points extraction. The method based idea auto-encoding wanted latent space while discarding non-essential information (and effectively preserving interpretability). interpretable representation (the bottleneck containing nothing but key-points) achieved by two-step regularization approach. first step evaluates transport distance from given set some average value barycenter Wasserstein distance). second controls deviations applying random geometric deformations synchronously initial and encoded landmarks. We demonstrate effectiveness both semi-supervised training scenarios using 300-W, CelebA, MAFL datasets. proposed paradigm shown prevent overfitting, detection quality improve beyond state-of-the-art models.