DATA.md

DATASET Preparation

This section provides instructions on downloading and preparing the following datasets:

• ShapeSplat Dataset
• ModelSplat Dataset
• ShapeSplat-Part Dataset

ShapeSplat Dataset

1. Request Access
   Obtain access to the ShapeNet dataset from the Hugging Face page. Approval may take approximately 1–2 days.

2. Download the Dataset
   Once access is granted, you can download the dataset using the following steps:

   mkdir -p gs_data/shapesplat
   cd gs_data/shapesplat
   
   huggingface-cli download ShapeNet/ShapeSplatsV1 --token <Your Hugging Face Token> --repo-type dataset --local-dir .

   Replace <Your Hugging Face Token> with your personal access token from Hugging Face.

3. Process the Dataset
   After downloading the dataset, you will find multiple .zip files in the target folder. Use the provided script ./scripts/unzip_shapesplat.sh to extract all the files. Update the script with the following values:

   • <SOURCE_DIR>: Path to the folder where the downloaded .zip files are located.
   • <DEST_DIR>: Path to the folder where you want the data to be extracted.

   The extracted files will follow the structure below, with all .ply files named in the format: < {category_id}-{object_id}.ply>.

   shapesplat_ply
   ├── 04256520-81d3c178c611199831e05c4a367a9ebd.ply
   ├── 03001627-d388617a719f23a191d0dff7aea42471.ply
   ├── 03001627-2060f4d31feb73dd7762ba3a489df006.ply
   ├── 03991062-ad258c77538632867bc41009043d88b0.ply
   ├── 03691459-707f0e44e935dd55edfd593a4f114036.ply
   
   

4. Prepare the Dataset config
   The ShapeNet55GS.yaml file is located in the cfgs/dataset_configs directory. Please set the following paths:

   • DATA_PATH: Set this to datasets/shapenet_split.
   • GS_PATH: Set this to the directory where you extracted the .ply files.

ModelSplat Dataset

1. Download the Dataset
   Similarly, you can download the modelsplat in ModelNetSplats Dataset Release

   mkdir -p gs_data/modelsplat
   cd gs_data/modelsplat
   
   huggingface-cli download ShapeSplats/ModelNet_Splats --token  <Your Hugging Face Token> --repo-type dataset --local-dir .

   Replace <Your Hugging Face Token> with your personal access token from Hugging Face.

2. Process the Dataset
   After downloading there will be a lot of .zip files, we want to unzip to following data structure:

   modelsplat
   ├── airplane
   │   ├── train 
   │   │   ├── airplane_0001
   │   │   │   ├── point_cloud.ply
   │   │   ├── airplane_0002
   │   │   │   ├── point_cloud.ply
   │   │   ├── .......
   │   ├── test                 
   │   │   ├── airplane_0627
   │   │   │   ├── point_cloud.ply
   │   │   ├── airplane_0002
   │   │   │   ├── airplane_0628.ply
   │   │   ├── .......
   ├── bathtub
   ├── .......
   
   

   Please use the provided script ./scripts/unzip_modelsplat.sh and configure the paths for unzipping.

   • <SOURCE_DIR>: Path to the folder where the downloaded .zip files are located.
   • <DEST_DIR>: Path to the folder where you want the data to be extracted.

3. Prepare the Dataset config
   The ModelNet10GS.yaml and ModelNet40GS.yaml files are located in the cfgs/dataset_configs directory. Please set the following paths:

   • DATA_PATH: Set this to datasets/modelnet_split folder inside the repo.
   • GS_PATH: Set this to the directory where you extracted the .ply files.

ShapeSplat-Part Dataset

1. Download the Dataset
   You can officially download the Shape-Part annotations from the following link. If the official link is unavailable, you can also download the dataset from our Hugging Face repository.

   After unzipping shapenetcore_partanno_segmentation_benchmark_v0_normal.zip, you need to update the following environment variables, more details please forward to segmentation_gs:

   • PARTANNO_ROOT
   • GS_ROOT (Note: This should be the same as the GS_PATH used for ShapeSplat)
   • PC_ROOT (Note: This is the path where to shape_data the pointcloud annotation)