A U-Net Architecture for Inpainting Lightstage Normal Maps

In this paper, we investigate the inpainting of normal maps that were captured from a lightstage. Occlusion of parts of the face during performance capture can be caused by the movement of, e.g., arms, hair, or props. Inpainting is the process of interpolating missing areas of an image with plausible data. We build on previous works about general image inpainting that use generative adversarial networks (GANs). We extend our previous work on normal map inpainting to use a U-Net structured generator network. Our method takes into account the nature of the normal map data and so requires modification of the loss function. We use a cosine loss rather than the more common mean squared error loss when training the generator. Due to the small amount of training data available, even when using synthetic datasets, we require significant augmentation, which also needs to take account of the particular nature of the input data. Image flipping and inplane rotations need to properly flip and rotate the normal vectors. During training, we monitor key performance metrics including the average loss, structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR) of the generator, alongside the average loss and accuracy of the discriminator. Our analysis reveals that the proposed model generates high-quality, realistic inpainted normal maps, demonstrating the potential for application to performance capture. The results of this investigation provide a baseline on which future researchers can build with more advanced networks and comparison with inpainting of the source images used to generate the normal maps.