Processing the 3D Heritage Data Samples Based on Combination of GNN and GAN

The heritage objects hold historical values because they are associated with culture and have a significant impact on society. 3D reconstruction of the heritage objects is an important research to preserve and promote cultural heritage. A lot of effort has been taken to digitalize the heritages in the researched communities. The state-of-the-art methods that are current on reconstructing the heritage objects based on geometric modeling and computer graphics. Additionally, the methods based on machine learning techniques are widely researched in recent years. However, the obtained results still need to be improved, depending on characteristics of input data. This paper proposes a method for denoising the 3D heritage objects from scanned data based on a combination of Graph Neural Network (GNN) and Generation Adversarial Network (GAN). Our method includes the following steps. We first collect and process data of the real heritage objects. After meshing the object surface, we enrich the data training process by adding a normal vector for each 3D point and face. The information of both positions and normal vector of points and faces can support the denoising process. In the next step, we create a GNN based on the U-Net architecture to extract the features of 3D objects. In order to increase the performance of denoising model, we combine with a GAN framework to generate samples, such that they are as close to ground truth as possible; until the discriminator cannot realize the real or fake objects. Comparing the existing methods, our proposed method obtained better results and is closer to the ground truth data.