Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A. Yeh, Rana Hanocka

The task of crafting procedural programs capable of generating structurally valid 3D shapes easily and intuitively remains an elusive goal in computer vision and graphics. Within the graphics community, generating procedural 3D models has shifted to using node graph systems. They allow the artist to create complex shapes and animations through visual programming. Being a high-level design tool, they made procedural 3D modeling more accessible. However, crafting those node graphs demands expertise and training. We present GeoCode, a novel framework designed to extend an existing node graph system and significantly lower the bar for the creation of new procedural 3D shape programs. Our approach meticulously balances expressiveness and generalization for part-based shapes. We propose a curated set of new geometric building blocks that are expressive and reusable across domains. We showcase three innovative and expressive programs developed through our technique and geometric building blocks. Our programs enforce intricate rules, empowering users to execute intuitive high-level parameter edits that seamlessly propagate throughout the entire shape at a lower level while maintaining its validity. To evaluate the user-friendliness of our geometric building blocks among non-experts, we conduct a user study that demonstrates their ease of use and highlights their applicability across diverse domains. Empirical evidence shows the superior accuracy of GeoCode in inferring and recovering 3D shapes compared to an existing competitor. Furthermore, our method demonstrates superior expressiveness compared to alternatives that utilize coarse primitives. Notably, we illustrate the ability to execute controllable local and global shape manipulations.

A demo video of our program is available on our project page.

• Python 3.8
• CUDA 11.8
• GPU, minimum 8 GB ram
• During training, a machine with 5 CPUs is recommended
• During visualization and sketch generation, we recommend a setup with multiple GPU nodes, refer to the additional information to run in parallel on all available nodes
• During test-set evaluation, generation of raw shapes for a new dataset, and during stability metric evaluation, a single node with 20 CPUs is recommended

Clone and create the Conda environment

git clone https://github.com/threedle/GeoCode.git
cd GeoCode
conda env create -f environment.yml
conda activate geocode
python setup.py install

# Install Blender 3.2 under `~/Blender`
(sudo) chmod +x ./scripts/install_blender3.2.sh
./scripts/install_blender3.2.sh

# Download the dataset (`~/datasets`), checkpoint (`~/models`) and blend file (`~/blends`) of the `chair` domain
python scripts/download_ds.py --domain chair --datasets-dir ~/datasets --models-dir ~/models --blends-dir ~/blends

vase and table domains are also available

Run the test for the chair domain using the downloaded checkpoint, make sure the directories match the directories that were used in the download_ds.py step

cd GeoCode
conda activate geocode
python geocode/geocode.py test --blender-exe ~/Blender/blender-3.2.0-linux-x64/blender --blend-file ~/blends/procedural_chair.blend --models-dir ~/models --dataset-dir ~/datasets/ChairDataset --input-type pc sketch --phase test --exp-name exp_geocode_chair

This will generate the results in the following directory structure, in

<datasets-dir>
│
└───ChairDataset
    │
    └───test
        │
        └───results_exp_geocode_chair
            │
            └───barplot                    <-- model accuracy graph
            └───obj_gt                     <-- 3D objects of the ground truth samples
            └───obj_predictions_pc         <-- 3D objects predicted from point cloud input
            └───obj_predictions_sketch     <-- 3D objects predicted from sketch input
            └───yml_gt                     <-- labels of the ground truth objects
            └───yml_predictions_pc         <-- labels of the objects predicted from point cloud input
            └───yml_predictions_sketch     <-- labels of the objects predicted from sketch input

We also provide a way to automatically render the resulting 3D objects. Please note that this step is GPU intensive due to rendering, the use of multiple nodes with GPU is recommended. Please see the additional information for running this in parallel.

cd GeoCode
conda activate geocode
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_chair.blend -b --python visualize_results/visualize.py -- --dataset-dir ~/datasets/ChairDataset --phase test --exp-name exp_geocode_chair

this will generate the following additional directories under results_exp_geocode_chair:

            ⋮
            └───render_gt                  <-- renders of the ground truth objects
            └───render_predictions_pc      <-- renders of the objects predicted from point cloud input
            └───render_predictions_sketch  <-- renders of the objects predicted from sketch input

Training from a checkpoint or new training is done similarly, and only depends on the existence of a latest.ckpt checkpoint file in the experiment directory (under ~/models in this example). Please note that training using our checkpoints will show a starting epoch of 0.

cd GeoCode
conda activate geocode
python geocode/geocode.py train --models-dir ~/models --dataset-dir ~/datasets/ChairDataset --nepoch=600 --batch_size=33 --input-type pc sketch --exp-name exp_geocode_chair

Open one of the Blend files using Blender 3.2.

To modify the shape using the parameters and to inspect the Geometry Nodes Program click the "Geometry Node" workspace at the top of the window

Then you will see the following screen

For logging during training, we encourage the use of neptune.ai. First open an account and create a project, create the file GeoCode/config/neptune_config.yml with the following content:

neptune:
  api_token: "<TOKEN>"
  project: "<POJECT_PATH>"

When downloading one or more domain using:

python scripts/download_ds.py --domain chair --datasets-dir ~/datasets --models-dir ~/models --blends-dir ~/blends
python scripts/download_ds.py --domain vase --datasets-dir ~/datasets --models-dir ~/models --blends-dir ~/blends
python scripts/download_ds.py --domain table --datasets-dir ~/datasets --models-dir ~/models --blends-dir ~/blends

The resulting directory structure will be (example for the chair domain):

<datasets-dir>
│
└───ChairDataset
    │
    └───recipe.yml
    │
    └───train
    │   └───obj_gt
    │   └───point_cloud_fps
    │   └───point_cloud_random
    │   └───sketches
    │   └───yml_gt
    │   └───yml_gt_normalized
    │
    └───val
    │   └───obj_gt
    │   └───...
    │
    └───test
        └───obj_gt
        └───...
        
<models-dir>
│
└───exp_geocode_chair
    │
    └───procedural_chair_epoch585_ckpt.zip
    └───procedural_chair_last_ckpt.zip
    └───procedural_chair_epoch585.ckpt
    └───last.ckpt

<blends-dir>
│
└───procedural_chair.blend

When visualizing the results, we render an image for each ground truth and prediction 3D object that were created while running the test-set evaluation. Since this is GPU intensive task, we provide a way to run this in parallel on multiple GPU machines. To do so, simply add the flags --parallel 10 --mod $NODE_ID to the visualization command, for example, for 10 nodes:

cd GeoCode
conda activate geocode
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_chair.blend -b --python visualize_results/visualize.py -- --dataset-dir ~/datasets/ChairDataset --phase test --exp-name exp_geocode_chair --parallel 10 --mod $NODE_ID

where $NODE_ID is the node id.

You can optionally edit the dataset_generation or the camera_angles sections in the appropriate recipe YAML file. For example, for the chair domain, edit the following recipe file: GeoCode/dataset_generator/recipe_files/chair_recipe.yml. We encourage the user to inspect the relevant Blend file before modifying the recipe file.

In this step no GPU and no conda env are required.

For example, generating the val, test, and train datasets for the chair domain, with 3, 3, and 30 shape variation per parameter value, is done using the following commands:

cd GeoCode
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_chair.blend -b --python dataset_generator/dataset_generator.py -- generate-dataset --dataset-dir ~/datasets/MyChairDataset --domain chair --phase val --num-variations 3 --parallel 20
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_chair.blend -b --python dataset_generator/dataset_generator.py -- generate-dataset --dataset-dir ~/datasets/MyChairDataset --domain chair --phase test --num-variations 3 --parallel 20
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_chair.blend -b --python dataset_generator/dataset_generator.py -- generate-dataset --dataset-dir ~/datasets/MyChairDataset --domain chair --phase train --num-variations 30 --parallel 20

For vase domain

~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_vase.blend -b --python dataset_generator/dataset_generator.py -- generate-dataset --dataset-dir ~/datasets/MyVaseDataset --domain vase --phase val --num-variations 3 --parallel 20
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_vase.blend -b --python dataset_generator/dataset_generator.py -- generate-dataset --dataset-dir ~/datasets/MyVaseDataset --domain vase --phase test --num-variations 3 --parallel 20
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_vase.blend -b --python dataset_generator/dataset_generator.py -- generate-dataset --dataset-dir ~/datasets/MyVaseDataset --domain vase --phase train --num-variations 30 --parallel 20

For table dataset

~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_table.blend -b --python dataset_generator/dataset_generator.py -- generate-dataset --dataset-dir ~/datasets/MyTableDataset --domain table --phase val --num-variations 3 --parallel 20
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_table.blend -b --python dataset_generator/dataset_generator.py -- generate-dataset --dataset-dir ~/datasets/MyTableDataset --domain table --phase test --num-variations 3 --parallel 20
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_table.blend -b --python dataset_generator/dataset_generator.py -- generate-dataset --dataset-dir ~/datasets/MyTableDataset --domain table --phase train --num-variations 30 --parallel 20

Please note that the shapes generated in this step are already normalized.

This step does not require a conda env but requires GPU(s)

cd GeoCode
~/Blender/blender-3.2.0-linux-x64/blender ~/blends/procedural_vase.blend -b --python dataset_generators/sketch_generator.py -- --dataset-dir ~/datasets/MyChairDataset --phases val test train

You can also run this in parallel, for example, with 10 processes, by adding the flags --parallel 10 --mod $NODE_ID

Once you train on the new dataset, a preprocessing step of the dataset will be performed. The point cloud sampling and normalized label directories and files will be created under the dataset directory. Refer to the Download the datasets section at the top of this readme for the directory structure.

To evaluate a tested dataset phase using our stability metric use the following command:

cd GeoCode/stability_metric
python stability_parallel.py --blender-exe ~/Blender/blender-3.2.0-linux-x64/blender --dir-path ~/datasets/ChairDataset/val/obj_gt

In our tests we also used COSEG dataset. A preprocessing step is required when working with external datasets.

• 3D shapes should be in .obj format
• 3D shapes should be normalized before training or testing

we provide a command to download an additional script that allows working with COSEG samples. Additionally, we provide the code to simplify a dataset (mesh simplification). Please refer to the README.md file that is downloaded along with the scripts after executing the following command:

cd GeoCode
scripts/download_ds_processing_scripts.py
~/Blender/blender-3.2.0-linux-x64/blender -b --python dataset_processing/prepare_coseg.py -- --help

• common - util packages and classes that are shared by multiple other directories
• config - contains the neptune.ai configurations file
• data - dataset related classes
• dataset_processing - scripts that are intended to manipulate existing datasets
• geocode - main training and testing code
• models - point cloud and sketch encoders and the decoders network
• scripts - contains scripts to set up our datasets and saved checkpoints and installing Blender
• stability_metric - scripts to evaluate a tested phase using our stability metric
• visualize_results - script to generate the renders for all ground truth and predicted shapes

@article{pearl2022geocode,
  title={GeoCode: Interpretable Shape Programs},
  author={Pearl, Ofek and Lang, Itai and Hu, Yuhua and Yeh, Raymond A. and Hanocka, Rana},
  booktitle={arXiv preprint arxiv:2212.11715},
  year={2022}
}