DNF-Net: a Deep Normal Filtering Network for Mesh Denoising

by Xianzhi Li, Ruihui Li, Lei Zhu, Chi-Wing Fu, and Pheng-Ann Heng.

Introduction

This repository is for our IEEE Transactions on Visualization and Computer Graphics (TVCG) 2020 paper 'DNF-Net: a Deep Normal Filtering Network for Mesh Denoising'. In this paper, we present a deep normal filtering network, called DNF-Net, for mesh denoising. To better capture local geometry, our network processes the mesh in terms of local patches extracted from the mesh. Overall, DNF-Net is an end-to-end network that takes patches of facet normals as inputs and directly outputs the corresponding denoised facet normals of patches. In this way, we can reconstruct the geometry from the denoised normals with feature preservation. Besides the overall network architecture, our contributions include a novel multi-scale feature embedding unit, a residual learning strategy to remove noise, and a deeply-supervised joint loss function.

Usage

Environment Requirements

• Python 2.7
• Tensorflow 1.9.0
• Ubuntu 16.04.6
• CUDA 10.2

Try our previously-trained network:

(1) Step 1: prepare network input
For your convenience, we provided the three noisy meshes shown in Figure 8 of our paper. You can download them here. Since our DNF-Net takes patches as network inputs, we directly provide the patches of each mesh; see the three h5 files in network-inputs folder. You can directly use them as network inputs.

Or you can also use your own noisy meshes. You need to cut the noisy mesh into patches. Please download the C++ code here. Unzip the file, inside the Meshviewer folder, open meshviewer.cpp, and you will see the main function. In the prepareTestData() function, you need to specify the path and name of your noisy meshes. Then the function will generate patches per face automatically. Finally, use the code patch_to_h5.py inside the code folder (see this GitHub page) to generate h5 file.

(2) Step 2: testing to generate the denoised facet normals
Feed the h5 file to our trained network for testing. The python code is provided in the code folder --> test.py. After running, the network will output a txt file. In the txt file, the three values in each line indicate the denoised facet normal of each face.

(3) Step 3: generate the denoised mesh
Still use the C++ code downloaded in Step 1 to reconstruct the denoised mesh from the denoised facet normals. In the main function, there is a function called generateDenoisedMesh(). You need to specify the noisy mesh and the vertex iteration number. You can also give the path of ground-truth mesh to calculate the denoising error. For our provided three meshes, we have specified the corresponding recommended vertex iteration number.

Re-train our network:

If you want to re-train our network, we provide the synthetic training dataset here. The python code is provided in the code folder --> train.py.

Questions

Please contact 'lixianzhi123@gmail.com'