Recognizing Tiny Faces

Objects are naturally captured over a continuous range of distances, causing dramatic changes in appearance, especially at low resolutions. Recognizing such small objects at range is an open challenge in object recognition. In this paper, we explore solutions to this problem by tackling the fine-grained task of face recognition. State-of-the-art embeddings aim to be scale-invariant by extracting representations in a canonical coordinate frame (by resizing a face window to a resolution of say, 224x224 pixels). However, it is well known in the psychophysics literature that human vision is decidedly scale variant: humans are much less accurate at lower resolutions. Motivated by this, we explore scale-variant multiresolution embeddings that explicitly disentangle factors of variation across resolution and scale. Importantly, multiresolution embeddings can adapt in size and complexity to the resolution of input image on-the-fly (e.g., high resolution input images produce more detailed representations that result in better recognition performance). Compared to state-of-the-art "one-size-fits-all" approaches, our embeddings dramatically reduce error for small faces by at least 70% on standard benchmarks (i.e. IJBC, LFW and MegaFace).