GANerated Hands for Real-Time 3D Hand Tracking from Monocular RGB

TLDR

The study aims to enable real‑time 3D hand tracking from a single RGB camera by introducing a novel synthetic data generation pipeline. The method trains a CNN to predict hand pose, guided by a kinematic 3D hand model, using a geometry‑aware image‑to‑image translation network that maps synthetic images to realistic ones via adversarial, cycle‑consistency, and geometric consistency losses. The system achieves superior performance over existing methods on challenging RGB‑only hand‑tracking benchmarks.

Abstract

We address the highly challenging problem of real-time 3D hand tracking based on a monocular RGB-only sequence. Our tracking method combines a convolutional neural network with a kinematic 3D hand model, such that it generalizes well to unseen data, is robust to occlusions and varying camera viewpoints, and leads to anatomically plausible as well as temporally smooth hand motions. For training our CNN we propose a novel approach for the synthetic generation of training data that is based on a geometrically consistent image-to-image translation network. To be more specific, we use a neural network that translates synthetic images to "real" images, such that the so-generated images follow the same statistical distribution as real-world hand images. For training this translation network we combine an adversarial loss and a cycle-consistency loss with a geometric consistency loss in order to preserve geometric properties (such as hand pose) during translation. We demonstrate that our hand tracking system outperforms the current state-of-the-art on challenging RGB-only footage.