|
Three-dimensional (3D) model with high-quality texture is a powerful way to interact between virtual and reality 3D scene. It has significant applications in digital preservation of cultural heritage, virtual reality / augmented reality (VR/AR), medical imaging and other domains. High-quality texture reconstruction is one of the essential elements for 3D model. Due to the camera pose error and inaccurate model, current texture reconstruction methods could not eliminate the artifacts like blurring, ghosting and discontinuity.
In this work, multi-view high-definition images are captured for texture mapping. Normal-weight, depth-weight and edge-weight parameter are introduced to evaluate texture color confidence, respectively. The normalized weight factors are multiplied to generate a comprehensive weight parameter. By taking a weighted average of projected texture images, discontinuity of texture can be smoothed to a large extent. For large misalignment, bidirectional similarity (BDS) function, which represent the structural similarity between two images, are utilized to improve the texture image. The energy function is composed of two terms. One is Euclidean distance between target texture image and merged texture image. The other is BDS between target texture image and original texture image. By minimizing the energy function, the texture image could generate small local displacement while retaining the original structural information. The target and merged images are optimized alternately, the target image is calculated by patch-match algorithm, and the merged image is derived from weighted average of target images. The method we proposed could produce seamless texture comparing with Markov random field (MRF) algorithm. The definition could be higher than the camera parameter optimization algorithm. The experimental results show that structural similarity (SSIM) between the reconstructed images and the ground truth is higher than traditional algorithms. When dealing with inaccurate models with less than 10,000 facets, the SSIM value could be doubled compared with current algorithms.
|
