ShapeNetSDF

Signed Distance Field (SDF) point samples derived from ShapeNet Core, for training and evaluating implicit neural representations / neural fields on 3D shapes.

Each shape is converted into a watertight manifold, normalized into the unit cube [-1, 1]³, and sampled into three point sets (uniform, surface, groundtruth), each stored as a [N, 4] float32 array of [x, y, z, sdf].

This dataset is produced by process_shapenet_to_sdf.py from the wsr.pytorch neural-field codebase.

Dataset structure

The dataset is stored as sharded Parquet (one config per category, with train / val / test splits). Each row is one shape; the three point sets (uniform, surface, groundtruth) are stored as fixed-shape [262144, 4] float32 tensors in their own columns.

ShapeNetSDF/
├── data/                           # heavy data: sharded parquet, one folder per category
│   └── <category>/                 # 50 category configs (chair, table, airplane, ...)
│       ├── train-NNNNN-of-NNNNN.parquet   # rows: model_id, uniform, surface, groundtruth
│       ├── val-NNNNN-of-NNNNN.parquet
│       └── test-NNNNN-of-NNNNN.parquet
├── meta/                           # lightweight metadata (split id lists, labels, subsets)
│   ├── <category>/                 # per-category split id lists
│   │   └── train.txt / val.txt / test.txt
│   └── all/                        # global splits + curated subsets + labels
│       ├── train.txt / val.txt / test.txt   # global splits (all categories)
│       ├── 10k10c.txt / 5k10c.txt / 5k5c.txt / 100_5c.txt   # curated subsets (model-id lists)
│       └── labels.json             # category ↔ model_id mappings
└── README.md

• Each Parquet row has columns model_id (string) and uniform / surface / groundtruth, each an Array2D[262144, 4] float32 tensor ([x, y, z, sdf] per point — see File format below).

• The all config globs every category's Parquet, so load_dataset(..., "all") streams all 50 categories together. The split id lists under meta/ (and curated subsets) remain for selecting specific model ids.

• 50 categories: airplane, bag, basket, bathtub, bed, bench, birdhouse, bookshelf, bottle, bowl, bus, cabinet, camera, can, cap, car, chair, dishwasher, display, earphone, faucet, file_cabinet, guitar, jar, keyboard, knife, lamp, laptop, loudspeaker, mailbox, microphone, motorbike, mug, piano, pillow, pistol, printer, remote, rifle, rocket, skateboard, sofa, stove, table, telephone, tower, train, trash_bin, washer, watercraft (plus the aggregated all/ folder).

• Total size: ~512 GB.

File format

Each point set (uniform / surface / groundtruth) is a float32 array of shape [262144, 4]:

column	meaning	range
0 (x)	x coordinate	[-1, 1]
1 (y)	y coordinate	[-1, 1]
2 (z)	z coordinate	[-1, 1]
3 (sdf)	signed distance to the surface	negative inside, positive outside

• uniform — N = 64³ = 262,144 points sampled uniformly in [-1, 1]³, with their exact signed distance. Use these to supervise the field in free space.
• surface — points sampled on the mesh surface with a small Gaussian perturbation (noise scale ≈ 0.02), so SDF values are close to (but not exactly) zero. Intended for training near the surface.
• groundtruth — exact on-surface points (SDF ≈ 0), with no noise. Intended for evaluation (e.g. surface reconstruction / Chamfer / IoU).

Splits

Each category config has train / val / test splits (~80% / 10% / 10%, seed 42), encoded as the Parquet partitions. Plain-text model_id lists under meta/ remain for global splits and curated subsets:

• meta/all/train.txt / val.txt / test.txt — global splits over all categories (per-category lists live in meta/<category>/).
• meta/all/{10k10c,5k10c,5k5c,100_5c}.txt — curated subsets (<num_models><num_categories>c, e.g. 5k5c = 5,000 models across 5 categories) for quick experiments. Filter the loaded dataset by model_id.
• meta/all/labels.json — {"category_to_filename": {...}, "filename_to_category": {...}} mapping each category to its model ids and vice versa.

Usage

from datasets import load_dataset

# One category, one split. Array2D columns decode to numpy [262144, 4] arrays.
ds = load_dataset("EPFL-IVRL/ShapeNetSDF", "chair", split="train").with_format("numpy")
row = ds[0]
xyz, sdf = row["uniform"][:, :3], row["uniform"][:, 3]   # [262144, 3], [262144]
# row also has "surface" and "groundtruth", same shape, plus "model_id".

# All categories together (streamed to avoid downloading everything at once):
ds_all = load_dataset("EPFL-IVRL/ShapeNetSDF", "all", split="train", streaming=True)

Or download the Parquet files for a category locally:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="EPFL-IVRL/ShapeNetSDF",
    repo_type="dataset",
    allow_patterns=["data/chair/**", "meta/all/labels.json"],
    local_dir="ShapeNetSDF",
)

How it was created

For each ShapeNet Core model (models/model_normalized.obj):

1. Convert the mesh to a watertight manifold (via point_cloud_utils).
2. Normalize into the unit cube centered at the origin (scale 0.98 of the max vertex distance).
3. Sample uniform, surface (+noise), and groundtruth point sets.
4. Compute the signed distance for every point and save as [N, 4] float32.

Processing is deterministic (per-model_id seeding), so results are reproducible. See neural_field/scripts/process_shapenet_to_sdf.py for the full pipeline and configuration.

License & citation

This dataset is created as part of the CVPR 2026 Paper "Weight Space Representation Learning via Neural Field Adaptation".

@inproceedings{yang2026wsr,
  title     = {Weight Space Representation Learning via Neural Field Adaptation},
  author    = {Yang, Zhuoqian and Salzmann, Mathieu and S{\"u}sstrunk, Sabine},
  booktitle = {Proceedings of the IEEE/CVF Conference on
               Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

This dataset is derived from ShapeNet and is subject to the ShapeNet terms of use. Use it for non-commercial research only, and cite ShapeNet:

@article{chang2015shapenet,
  title   = {ShapeNet: An Information-Rich 3D Model Repository},
  author  = {Chang, Angel X. and Funkhouser, Thomas and Guibas, Leonidas and
             Hanrahan, Pat and Huang, Qixing and Li, Zimo and Savarese, Silvio
             and Savva, Manolis and Song, Shuran and Su, Hao and Xiao, Jianxiong
             and Yi, Li and Yu, Fisher},
  journal = {arXiv preprint arXiv:1512.03012},
  year    = {2015}
}