BRULÈ: Barycenter-Regularized Unsupervised Landmark Extraction
作者：
Highlights：
• The first method that predicts interpretable landmarks in unsupervised way.
• Unlike pre-trained models which require large datasets for pre-training their auto-encoders, our method needs just a dozen of images to compute barycenter.
• In a semi-supervised scenario, our method outperforms state-of-the-art models.
• Two types of regularization (barycenter and geometric transforms) are shown to suffice for auto-encoder to produce the image landmarks in the bottleneck.
• New type of cyclic/conditional GAN architecture that performs training with only one domain data and decomposes images into landmarks and style.
摘要
•The first method that predicts interpretable landmarks in unsupervised way.•Unlike pre-trained models which require large datasets for pre-training their auto-encoders, our method needs just a dozen of images to compute barycenter.•In a semi-supervised scenario, our method outperforms state-of-the-art models.•Two types of regularization (barycenter and geometric transforms) are shown to suffice for auto-encoder to produce the image landmarks in the bottleneck.•New type of cyclic/conditional GAN architecture that performs training with only one domain data and decomposes images into landmarks and style.
论文关键词：
论文评审过程：Received 5 August 2021, Revised 14 March 2022, Accepted 26 May 2022, Available online 27 May 2022, Version of Record 2 June 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2022.108816
