Neural Wavelet-domain Diffusion for 3D Shape Generation论文阅读

Neural Wavelet-domain Diffusion for 3D Shape Generation

This paper presents a new approach for 3D shape generation, enabling direct generative modeling on a continuous implicit representation in wavelet domain.

http://arxiv.org/abs/2209.08725

Abstract

• Contribution
1. A compact wavelet representation (i.e., a pair of coarse and detail coefficient volumes) based on biorthogonal wavelets and truncated signed distance field to implicitly encode 3D shapes.
2. A generator network formulated based on the diffusion probabilistic model.
3. A detail predictor network, formulated to produce compatible detail coefficients to enhance the fine details in the generated shapes.

Method

Data Preparation

1. Sample a signed distance field (SDF) and truncate its distance values to avoid redundant information.
💡
We truncate the distance values in the SDF to [−0.1, +0.1], so regions not close to object surface are clipped to a constant. This help to reduce the shape representation redundancy and enable the shape learning process to better focus on the shape’s structures and fine details.

2. Transform the truncated SDF to the wavelet domain to produce a series of multi-scale coefficient volumes.

Shape Learning

1. Train a pair of neural networks to learn the 3D shape distribution from the coarse and detail coefficient volumes.
we adopt the denoising diffusion probabilistic model to formulate and train the generator network to learn to iteratively refine a random noise sample for generating diverse 3D shapes in the form of the coarse coefficient volume.

2. Train a detail predictor network to learn to produce the detail coefficient volume from the coarse coefficient volume.

Shape generation

1. Starting from a random Gaussian noise sample, we first use the trained generator to produce the coarse coefficient volume then the detail predictor to produce an associated detail coefficient volume.

2. Perform an inverse wavelet transform.

Experiments