Learning Illumination from a Limited Field-of-View Image
Yuke Sun, Dan Li, Shuang Liu, Tian-Chi Cao, Yingsong Hu · 2020
Illumination estimation is a crucial part of augmented reality since it can make the virtual object look more realistic. However, single image-based lighting estimation is challenging due to the limited information. Here we combine deep learning with the spherical harmonic (SH) lighting which is widely used in precomputed radiance transfer. Specifically, a convolutional neural network that predicts SH coefficients from an image is designed, trained and tested. Moreover, we construct a new dataset for training SH coefficients based on the existing panorama dataset. The method in this work can finally predict realistic lighting from a single, limited field-of-view image, and it presents better results in some cases compared with previous research.