Multi-modal data generation with a deep metric variational autoencoder

We present a deep metric variational autoencoder for multi-modal data generation. The employs triplet loss in the latent space, which allows conditional generation by sampling new embeddings space within each class cluster. approach is evaluated on dataset consisting of otoscopy images tympanic membrane with corresponding wideband tympanometry measurements. modalities this are correlated, as they represent different aspects state middle ear, but do not direct pixel-to-pixel correlation. shows promising results pairs and tympanograms, will allow efficient augmentation from sources.