Texture Atlas Compression Based on Repeated Content Removal

Abstract

Optimizing the memory footprint of 3D models can have a major impact on the user experiences during real-time rendering and streaming visualization, where the major memory overhead lies in high-resolution texture data. In this work, we propose a robust and automatic pipeline to content-aware, lossy compression for texture atlas. The design of our solution lies in two observations: 1) mapping multiple surface patches to the same texture region is seamlessly compatible with the standard rendering pipeline, requiring no decompression before any usage; 2) a texture image has background regions and salient structural features, which can be handled separately to achieve a high compression rate. Accordingly, our method contains three phases. We first factor out redundant salient texture contents by detecting such regions and mapping their corresponding 3D surface patches to a single UV patch via a UV-preserving re-meshing procedure. We then compress redundant background content by clustering triangles into groups by their color. Finally, we create a new UV atlas with all repetitive texture contents removed and bake a new texture via differentiable rendering to remove potential inter-patch artifacts. To evaluate the efficacy of our approach, we batch-processed a dataset containing 100 models collected online. On average, our method achieves a texture atlas compression ratio of 81.80% with an averaged PSNR and MS-SSIM scores of 40.68 and 0.97, a marginal error in visual appearance.

[Paper] [Vid] [Supp_mat]

Installation

Requirements:

• Microsoft Visual Studio 2019+ with Microsoft Visual C++ installed
• Cuda 10.2+
• Pytorch 1.6+
• OpenCV

Tested in Anaconda3 with Python 3.8 and PyTorch 1.12 + Cuda 11.3

Setup (Windows)

1. Install Microsoft Visual Studio 2019+ with Microsoft Visual C++.
2. Place the opencv_world455.dll file in the Code folder directory.
3. Install Cuda 10.2 or above. Note: Install CUDA toolkit from https://developer.nvidia.com/cuda-toolkit (not through anaconda)
4. Install the appropriate version of PyTorch compatible with the installed Cuda toolkit.
5. Install nvdiffrast in conda env. Follow the installation instructions.
6. Install Segment Anything:

   pip install git+https://github.com/facebookresearch/segment-anything.git

7. download the ViT-L SAM model, put it in the Code folder like Code\sam_vit_l_0b3195.pth

Examples

Additional test models can be downloaded here Unzip and place the model and texture file in the project input_Path folder:

e.g., input_Path\model.obj，input_Path\texture.png (Texture size is currently supported only for 1024x1024 dimensions.)

Change into the AtlasCompress\Code directory and run:

python AtlasCompress.py --PSNR_thred 38.0 --error_theta 10.0 --data_Path "input_path" --output_Path "output_path"

The results will be stored in the output_Path folder.

Cite

@article{luo23atlascompress,
   title = {Texture Atlas Compression Based on Repeated Content Removal},
   author = {Luo, Yuzhe and Jin, Xiaogang and Pan, Zherong and Wu, Kui and Kou, Qilong and Yang, Xiajun and Gao, Xifeng},
   booktitle = {SIGGRAPH Asia 2023 Conference Papers},
   pages = {1--10},
   year = {2023}
  }