Direct pose estimation from RGB images using 3D objects

We present a real-time monocular camera pose estimation algorithm for augmented reality applications. Proposed model is a small convolutional neural network that is trained to directly estimate 6 Degree of Freedom (6-DOF) camera pose from an RGB image. Our model is designed to run on real-time devices with low memory and computation power. Our model can estimate the camera pose in less than 1ms while keeping accuracy comparable to the state-of-the art. This was made possible by employing geometrically sound loss functions and algebraic constraints. Furthermore, we introduce a new synthetic dataset for demonstrating the proposed methods capabilities.