Learning Highlight Separation of Real High Resolution Portrait Image

This work presents an approach for highlight separation of real high resolution portrait image. In order to obtain reliable ground truth of real images, a controllable portrait image collection system with 156 groups of light sources has been built. It has 4 cameras to collect the portrait images of 36 subjects from different angles, and then we use 4 data processing strategies on these images to obtain 4 training datasets. Based on these datasets, 4 U-Net networks are trained by using a single image as input. To test and evaluate, we input the 2560∗2560 resolution images into 4 models, and finally determine the best data processing strategy and trained network. Our method creates precise and believable highlight separation results for 2560∗2560 high resolution images, including when the subject is not looking straight at the camera.