3D Face Reconstruction from Low-Resolution Images with Convolutional Neural Networks

During the past years, convolutional neural networks (CNNs) have widely spread as a powerful tool for tackling a variety of challenges posed in computer vision. Consequently, the trend neither does stop at 3D face reconstruction: Recently, several CNN-based approaches for reconstructing the dense 3D geometry of a face from only a single image have been introduced. However, while all of these methods deal with 3D face reconstruction in the high-resolution (HR) case, reconstruction in low-resolution (LR) surveillance scenarios by means of CNNs has not received any attention so far. With this work, we address that gap, being the first to propose a CNN architecture specifically tailored to LR 3D face reconstruction: We introduce an end-to-end trainable CNN capable of simultaneously estimating 3D geometry and pose of a face given a single LR image. By coupling our network with a state-of-the-art LR face detector, we build a 3D face reconstruction pipeline ready for integration into real-world applications. We conduct systematic evaluation on LR versions of the in-the-wild AFLW2000-3D dataset, considering decreasing interocular distances (IODs) down to three pixels. The results show superior performance of the proposed method in the LR domain over state-of-the-art approaches, for both 3D face reconstruction and the closely related face alignment task.