In this paper, Auto Encoder and Variational Auto Encoder were compared in generating UV Position Map, which is one of the important factors for 3D face reconstruction. Both models were trained from the same MNIST data, and as a result of training, the...
In this paper, Auto Encoder and Variational Auto Encoder were compared in generating UV Position Map, which is one of the important factors for 3D face reconstruction. Both models were trained from the same MNIST data, and as a result of training, the performance of Variational Auto Encoder was better. This seems to be the effect of the reparameterization trick that Auto Encoder does not have. Since the encoder extracts the mean and variance of the input data and uses them, the decoder knows the distribution information of the input data, so more sophisticated images can be created. Through this, by using the flow field of the continuous UV position map generated by VAE, it can be added as a new input to NeRF, and a novel view with more natural and various angles can be created than that of the existing NeRF.