StyleMesh (CVPR2022)

This is the official repository that contains source code for the CVPR2022 paper StyleMesh.

[Arxiv] [Project Page] [Video]

If you find StyleMesh useful for your work please cite:

@inproceedings{hollein2022stylemesh,
  title={StyleMesh: Style Transfer for Indoor 3D Scene Reconstructions},
  author={H{\"o}llein, Lukas and Johnson, Justin and Nie{\ss}ner, Matthias},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6198--6208},
  year={2022}
}

Preprocessing

The following steps are necessary to prepare all data.

Data Setup

• Download Matterport and Scannet datasets (we use train scannet scenes only, see scripts/scannet/scannet_train.txt for a complete list)
• Prepare Scannet Scenes: see scripts/scannet/prepare_data and scripts/scannet/filter/filter_blurry.py. This extracts RGB images and removes blurry frames.
• Create texture parametrization for each mesh (requires blender): see scripts/matterport/create_uvs.py and scripts/scannet/create_uvs.py. This saves a new version of each mesh file, that contains a uv texture parametrization. A sample output can be found here: https://drive.google.com/file/d/1x76zVka-Nkk87M0qZYwFZ51FUv6pVL8y/view?usp=sharing
• Compile Matterport and Scannet renderers: see scripts/matterport/render_uv and scripts/scannet/render_uv for more details.
• Render uv maps for Mattertport and Scannet: see scripts/matterport/render_uvs.py and scripts/scannet/render_uvs.py. This preprocessing step speeds up the optimization by precomputing the texture lookups for each viewpoint. A sample output can be found here: https://drive.google.com/file/d/1QCOQivACD03ICIlO-E6ivEcfkDMuKYlM/view?usp=sharing

Project Setup

• Download VGG Model: https://bethgelab.org/media/uploads/pytorch_models/vgg_conv.pth or https://drive.google.com/file/d/1IpTw7zOjWFMYXgt2zThPYIJ8c-DJw3xU/view?usp=sharing
• Change dataset and vgg-model paths in scripts/train/* to your local machine locations
• Create a conda environment and use it for all further executions: see requirements.txt

Texture Optimization

The following steps allow to optimize a texture for a specific scene/style. You can easily select scene/style by modifying the corresponding values in the scripts (--scene for ScanNet and additionally --region for Matterport). It also allows to fine-tune loss weights, if you want to experiment with your own settings.

All style images that are used in the main paper are listed in styles.

By default, run files (texture, hparams, logging) are saved in style-mesh/lightning_logs

• Optimize Scannet: see scripts/train/optimize_texture_scannet*.sh
• Optimize Matterport: see scripts/train/optimize_texture_matterport*.sh

The suffix "with_angle_and_depth" is used for comparisons in Fig. 4,5,6,7,8,9,11. The suffixes "only2D" and "with_angle" are used for ablation study in Fig. 7. The suffix "dip" is used for the DIP baseline in Fig. 4,5,6

Render optimized Texture

You can render images with Mipmapping and Shading with our OpenGL renderers for each dataset. Alternatively, you can use the generated texture files after each optimization together with the generated meshes and view the textured mesh in any mesh viewer (e.g. Meshlab or Blender).

• Renderer Scannet: see scripts/scannet/render_mipmap_scannet.py
• Renderer Matterport: see scripts/matterport/render_mipmap_matterport.py

Evaluate Reprojection Error

We use the file scripts/eval/eval_image_folders.py for calculation of the reprojection error (Tab. 1)

Evaluate Circle Metric

We use the file scripts/eval/measure_circles.py for calculation of the circle's metric (Tab. 2, Fig. 8)