Wavelet transform-assisted generative model for efficient 3d deep shape generation

Unsupervised deep learning has been widely employed to generate high-quality samples. While its potential for point clouds generation tasks has recently been demonstrated remarkable results by many works, the training stage is still computationally expensive and has a high usage of GPU memory. This paper introduces a novel strategy to improve the performance of the generative model for point clouds. This is achieved by learning wavelet priors, which utilizes a score-based generative model in the wavelet domain. Taking advantage of the multi-scale representation provided by wavelet transform, it is more efficient to learn the gradient field of the log density, which indicates the distribution of 3D points. Specifically, in the training phase, multiply groups of point cloud data consisting of wavelet coefficients are used as the input to train the network by utilizing denoising score matching. Shape is iteratively updated from coarse to fine by applying Langevin dynamics. Experiments demonstrated that our model also achieves state-of-the-art performance in point clouds generation and auto-encoding, training at faster speed and lower GPU memory.