Jiao PAN | Liang LI | Hiroshi YAMAGUCHI | Kyoko HASEGAWA | Thufail I FADJAR | Bra MANTARA | Satoshi TANAKA
Herein, we propose a method for three-dimensional(3D) reconstruction of cultural heritage based on deep learning, which we apply to the reliefs of the Buddhist temple heritage of Borobudur Temple, in Indonesia. Some parts of the Borobudur reliefs have been hidden by stone walls and are not visible following the reinforcements during the Dutch rule. Today, only gray-scale photos of those hidden parts are displayed in the Borobudur Museum. First, we reconstruct 3D point clouds of the hidden reliefs from these photos and predict the pixel-wise depth information for each of them using a deep neural network model. We then apply our stochastic point-based rendering mechanism to produce a high-quality visualization of the reconstructed point clouds. We have achieved promising visualization results that provide us with an intuitive understanding of the valuable relief heritage that is no longer visible to ordinary visitors.