Denoising of Multiview Images to 3D Point Cloud Visualization using VisualSfM

Preserving our diverse cultural heritage for the future relies heavily on advanced 2D to 3D technologies. These cutting-edge tools are invaluable for documenting, recovering, and presenting tangible and intangible aspects of our heritage. Through these technological innovations, we ensure that the richness of our cultural legacy is not only preserved but also presented dynamically and accessible for future generations. To preserve heritage scenes, 2D smartphone/DSLR and 3D LiDAR cameras can capture Multiview images and visualize 3D of a scene in MeshLab using Structure from Motion workflow. However, 3D LiDAR sensors consume more power and can only capture scenes up to a certain distance. To address these issues, 2D smartphone cameras/DSLR cameras can be used. However, smartphone 2D cameras tend to exhibit more noise due to their smaller aperture and sensor size as compared to DSLR cameras. The Multiview 2D images of Chandramouleshwara Temple in Hubli, a 900-year-old Temple dedicated to the Hindu God Shiva, were captured using a Samsung mobile phone. The synthetically generated noisy images are denoised using NLM and DnCNN algorithms. The denoised images and ground truth images are compared in terms of SSIM and PSNR as image restoration performance parameters. It is observed that as the noise level increases in images, the SSIM and PSNR values are decreasing. However, the results obtained by NLM algorithm is outperformed in terms of PSNR by 4.31dB and SSIM by 0.415 compared to pre-trained DnCNN model in denoising of noisy Multiview images. The denoised Multiview images are then passed through Structure from Motion (SfM) workflow for feature correspondence, camera intrinsic/extrinsic estimation, and depth estimation using triangulation, followed by bundle adjustment to reduce reprojection error. The 3D point cloud of a denoised Multiview images is visualized in VisualSfM.