Occlusion-Robust FAU Recognition by Mining Latent Space of Masked Autoencoders

• DOI: 10.48550/arxiv.2212.04029
• Publication Date: 2022-01-01
• Authors: Minyang Jiang; Yongwei Wang; Martin J. McKeown; Z. Jane Wang
• Keywords: Autoencoder

Facial action units (FAUs) are critical for fine-grained facial expression analysis. Although FAU detection has been actively studied using ideally high quality images, it was not thoroughly under heavily occluded conditions. In this paper, we propose the first occlusion-robust recognition method to maintain performance heavy occlusions. Our novel approach takes advantage of rich information from latent space masked autoencoder (MAE) and transforms into features. Bypassing occlusion reconstruction step, our model efficiently extracts features faces by mining a pretrained autoencoder. Both node edge-level knowledge distillation also employed guide find mapping between vectors conditions, including random small patches large blocks, studied. Experimental results on BP4D DISFA datasets show that can achieve state-of-the-art performances occlusion, significantly outperforming existing baseline methods. particular, even proposed comparable as methods normal