摘要
A neural statistical approach to the reconstruction of novel viewpoint image us ing general regression neural networks(GRNN) is presented. Different color value will be obtained by watching the same surface point of an object from different viewpoints due to specular reflection, and the difference is related to the pos ition of viewpoint. The relationship between the position of viewpoint and the c olor of image is non linear, neural network is introduced to make curve fitting , where the inputs of neural network are only a few calibrated images with obvio us specular reflection. By training the neural network, network model is obtaine d. By inputing an arbitrary virtual viewpoint to the model, the image of the vir tual viewpoint can be computed. By using the method presented here, novel viewpo int image with photo realistic property can be obtained, especially images with obvious specular reflection can accurately be generated. The method is an image based rendering method, geometric model of the scene and position of lighting are not needed.
A neural statistical approach to the reconstruction of novel viewpoint image us ing general regression neural networks(GRNN) is presented. Different color value will be obtained by watching the same surface point of an object from different viewpoints due to specular reflection, and the difference is related to the pos ition of viewpoint. The relationship between the position of viewpoint and the c olor of image is non linear, neural network is introduced to make curve fitting , where the inputs of neural network are only a few calibrated images with obvio us specular reflection. By training the neural network, network model is obtaine d. By inputing an arbitrary virtual viewpoint to the model, the image of the vir tual viewpoint can be computed. By using the method presented here, novel viewpo int image with photo realistic property can be obtained, especially images with obvious specular reflection can accurately be generated. The method is an image based rendering method, geometric model of the scene and position of lighting are not needed.
基金
TianjinNaturalScienceFoundation(0 0 360 0 4 1 1 )
