ShapeNetPU

ShapeNetPU is a ShapeNet-derived dataset for point cloud upsampling, covering 13 object categories. It provides paired sparse and dense point clouds, multi-view training images, and test-set reference meshes.

💻 Code: ControlFlow3D · 📦 Pretrained weights: Google Drive

📄 Paper: arXiv:2610.21190 · NeurIPS 2026

Dataset summary · Download · Directory structure · Point groupings · Point cloud format · Use with ControlFlow3D · Citation

Examples from the training split: sparse input (2,048 points), dense ground truth (8,192 points), and eight rendered views per shape. Views are arranged left to right, top to bottom in each sample's renderings.txt order.

📰 News

• 🚀 2026-10-24: ControlFlow3D code and pretrained weights are released.
• 🗂️ 2026-10-07: ShapeNetPU is released.

📊 Dataset Summary

Property	Training split	Test split
Shapes	30,395	5,367
Object categories	13	13
Sparse input points per shape	2,048	2,048
Dense reference points per shape	8,192	8,192
Upsampling factor	4×	4×
Rendered views per shape	8	Not included
Reference meshes	Not included	OBJ

There are 35,762 shapes in total. Each shape is identified by its ShapeNet category ID and model ID. The supplied training and test splits have no shared category/model IDs; no separate validation split is provided.

🗂️ Categories and Splits

View category-level train/test counts

Synset ID	Category	Train	Test
02691156	airplane	2,705	478
02828884	bench	1,154	204
02933112	cabinet	1,128	200
02958343	car	6,085	1,074
03001627	chair	4,116	727
03211117	display	899	159
03636649	lamp	857	151
03691459	loudspeaker	1,114	197
04090263	rifle	1,933	341
04256520	sofa	2,578	455
04379243	table	5,729	1,011
04401088	telephone	868	153
04530566	watercraft	1,229	217
Total	13 categories	30,395	5,367

⬇️ Download

Install the Hub client and the dependencies for the minimal reading example, then download the dataset:

pip install -U huggingface_hub numpy plyfile
hf download BaldFatTiger/ShapeNetPU --repo-type dataset --local-dir data/ShapeNetPU

For the test split only, append --include "test/*" to the download command.

📁 Directory Structure

ShapeNetPU/
├── train/
│   └── <category_id>/
│       └── <model_id>/
│           ├── input/
│           │   ├── sparse_xyz.ply
│           │   └── rendering/
│           │       ├── <view_id>.png          # 8 selected views
│           │       ├── renderings.txt        # Ordered view filenames
│           │       └── rendering_metadata.txt
│           └── gt/
│               └── gt_xyz.ply
└── test/
    └── <category_id>/
        └── <model_id>/
            ├── input/
            │   └── sparse_xyz.ply
            └── gt/
                ├── gt_xyz.ply
                └── gt_mesh.obj

• <category_id> is a ShapeNet synset ID; <model_id> is the original model identifier.
• input/sparse_xyz.ply contains the sparse input cloud.
• gt/gt_xyz.ply contains the dense reference cloud.
• input/rendering/ is included in the training split only.
• gt/gt_mesh.obj is included in the test split only and can be used for mesh-based evaluation, such as point-to-surface distance.

🧩 Point Groupings

The same stored points support two complementary grouping views:

Grouping	Groups per shape	Definition	Overlap	Points per group: sparse / dense
Contiguous blocks	8	Equal-sized consecutive ranges in the original vertex order	No	256 / 1,024
Overlapping patches	10	Groups specified by the stored patch.vertex_indices lists	Yes	256 / 1,024

• Blocks are implicit: split the original vertex array into eight consecutive groups. Each point belongs to exactly one block; no separate block element is stored in the PLY.
• Patches are explicit: gather points using the zero-based index lists in the PLY patch element. Indices refer to vertices in the same file, and a point can belong to multiple patches.
• The loader names these modes part and patch, respectively. They are grouping conventions, not semantic part labels.

Preserve the original vertex order when constructing blocks. If vertices are reordered, carry their block memberships and remap patch indices to the new vertex positions.

📐 Point Cloud Format

Point clouds use binary little-endian PLY with the following fields:

Element	Property	Meaning
vertex	x, y, z	Float32 point coordinates
vertex	dense_index	Per-file integer indexing metadata; not a sparse-to-GT correspondence map
patch	center_x, center_y, center_z	Float32 patch-center coordinates
patch	vertex_indices	List of zero-based indices into this file's vertex array

The header comment patch_index_space local_vertex records this indexing convention.

Minimal Reading Example

After downloading, this example reads one sparse cloud and constructs both grouping views:

from pathlib import Path
import numpy as np
from plyfile import PlyData

sample_path = next(Path("data/ShapeNetPU/test").glob("*/*/input/sparse_xyz.ply"))
ply = PlyData.read(str(sample_path))
v = ply["vertex"].data
points = np.column_stack([v["x"], v["y"], v["z"]])
blocks = points.reshape(8, -1, 3)  # 8 contiguous, non-overlapping blocks
patches = [points[idx] for idx in ply["patch"].data["vertex_indices"]]

Use the corresponding gt/gt_xyz.ply file to read the dense reference cloud. Blocks have shape (8, 256, 3) for sparse inputs and (8, 1024, 3) for dense references; patches contains ten point arrays.

🖼️ Multi-View Renderings

Each training shape has eight 137 × 137 RGBA PNG images, selected from the source 24-view rendering set. Selected filenames range from 00.png to 23.png and can differ between shapes. Renderings are included in the training split only.

File	Contents
<view_id>.png	A selected rendered view
renderings.txt	The eight selected image filenames, in view order
rendering_metadata.txt	Eight rows of camera parameters, aligned with renderings.txt

The five metadata columns are azimuth, elevation, in-plane rotation, distance, and field of view, following the upstream rendering convention.

💡 Usage Notes

• Use the supplied train/test split. If validation data are needed, derive them from the training split.
• For point-cloud-only evaluation, use sparse test clouds as inputs; dense clouds and meshes are evaluation references.
• Apply consistent coordinate transforms to inputs, predictions, references, and meshes when computing distances, and report normalization and metric scaling.

🚀 Using ShapeNetPU with ControlFlow3D

The official repository provides the dataset-to-model workflow:

• Prepare point and image features.
• Choose source-file or memmap loading; memmap is optional and recommended.
• Follow the training and evaluation commands.

📝 Citation and Acknowledgments

If you find ShapeNetPU useful for your research, please consider citing our work as an acknowledgment:

@misc{liu2026controlflow3d,
  title={ControlFlow3D: Distilling Multi-View Knowledge into Latent Flow Matching for Point Cloud Upsampling},
  author={Yuang Liu and Zhi Zuo and Zhengkai Zhao and Lirui Zhang and Pan Gao and Hao Feng and Zhengzhe Liu},
  year={2026},
  eprint={2610.21190},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.21190},
}

Please also cite both ShapeNet and 3D-R2N2 to acknowledge the underlying 3D models and the source multi-view renderings, respectively.

ShapeNet — BibTeX from the official ShapeNet dataset card:

@techreport{shapenet2015,
  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},
  number      = {arXiv:1512.03012 [cs.GR]},
  institution = {Stanford University --- Princeton University --- Toyota Technological Institute at Chicago},
  year        = {2015}
}

3D-R2N2 — BibTeX from the authors' official repository:

@inproceedings{choy20163d,
  title={3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction},
  author={Choy, Christopher B and Xu, Danfei and Gwak, JunYoung and Chen, Kevin and Savarese, Silvio},
  booktitle = {Proceedings of the European Conference on Computer Vision ({ECCV})},
  year={2016}
}

We thank the contributors of ShapeNet and 3D-R2N2, as well as the original 3D model creators, for their open-source contributions and source assets.

📜 License and Terms of Use

ShapeNetPU is derived from ShapeNet. The underlying assets remain subject to the ShapeNet Terms of Use, which restrict use to non-commercial research and educational purposes, and to any applicable licenses of the original model sources.

Users must review and comply with the upstream terms before using the data. ShapeNet's terms permit sharing with research associates and colleagues only if they first agree to be bound by those terms. This dataset card does not grant unrestricted redistribution rights or additional rights to the underlying third-party assets.

📬 Contact

For questions about ShapeNetPU or to report a data issue, please open an issue in the ControlFlow3D repository. Include the split, category ID, and model ID when reporting a sample-specific issue.