Exposure-NeuS: Multi-View Head Reconstruction Using Exposure-Distorted Data
Zhixun Cheng, Bo Peng, Ning An · 2024
We propose Exposure-NeuS, a high fidelity human head reconstruction method from multi-view images, even from exposure-distorted images. Previous methods either did not reconstruct with sufficient accuracy, such as failing to reconstruct hair, or resulted in artifacts or holes in the reconstructed results due to exposure in the input data. To solve these challenging problems, we improve the original fine sampling of the model by using piecewise power function sampling to obtain sampling points closer to the real surface. And we develop a prior latent code to implicitly model factors such as environment and exposure. Experiments show that we eliminate or reduce the erroneous facial depressions and hair holes that occur in the baseline method during reconstruction. Released code is provided as follows:https://github.com/zhanliangOO/Exposure-NeuS.git