Image Based 3D Model Reconstruction of Ethiopian Museum Artifacts Using Deep Learning Technique

Abstract

Museum heritages are the legacy that have been passed from generation to generation. Presently, in the world, particularly in Ethiopia, several tangible heritage artifacts are available with great care physically. However, some artifacts were handled unprofessionally and are exposed to damages due to various reasons that include heritage damage due to lack of awareness or ignorance of the stakeholders, heritage theft, inappropriate conservation practices, and natural damage or deterioration. Moreover, the quality of services offered by these physical museum artifacts is not good at all for visitors. Therefore, to keep these artifacts safe and have global importance, it is important to digitally record, preserve and display these heritages. In this research, we have proposed an image-based 3D model reconstruction system for Ethiopian museum artifacts using a deep learning technique called NeRF with a single 2D image supervision only. The model has been trained and optimized to reconstruct the 3D model of Ethiopian museum artifacts with a marching cube algorithm in limited cost and time, in which visitors and researchers can be immersed and understand a culture and ancient civilization in the museums with the help of augmented and virtual reality applications. To the knowledge of these authors, there were no publicly available datasets of Ethiopian museum artifacts. Therefore, we started by collecting and preparing six different types of museum artifact datasets from the National Museum of Ethiopia and the Jimma museum. During experimental work, the effect of different batch sizes, learning rates, image resolution, and optimizers are tested and analyzed to fit the model and achieve a high-quality 3D model of the museum artifacts. In this study, differentiable volume rendering has been used for rendering tasks. The experiment is done on Google Colab pro with 30 epochs. We have measured the performance of a model using MSE and PSNR in our test cases, which contain a single 2D image. When analyzing the experimental result, we have achieved an effective 3D models for all collected Ethiopian museum at image resolution 504×378, a learning rate of 5×10- 4, batch size of 1024, and optimizer of a ranger.