Outdoor Illumination Estimation Via All Convolutional Neural Networks

When inserting a virtual object into an outdoor image, the recovery of sun orientation has a critical effect on the fusion of virtual objects and real scenes. For the problem of sun orientation estimation from a single outdoor image, a new end-to-end learning method is proposed in this paper. A luminance channel of sky region is introduced into the input to enhance the extracted image features. A network only consists of convolutional layers is designed to improve the ability of extracting image features. Pruning and quantization are used to compress the network, resulting in a large reduction in the number of network parameters and the storage space, only with a slight loss of precision.