3FabRec: Fast Few-Shot Face Alignment by Reconstruction

Figure 1: The framework of 3FabRec consisting of: (leftmost box) a first, unsupervised training stage that trains a low-dimensional latent space via an adversarial (generative) autoencoder on a large dataset of unlabeled faces, (middle box) subsequent supervised training with few annotated faces. The rightmost box shows results from testing the framework trained on only the 10 images of the middle box with original faces (top row), reconstructed faces via the autoencoder (middle row), and confidence heatmaps (bottom row).