PCOU3D Synthetic Database

[EN] English
PCOU3D (PointCloud Oulu 3D) is a synthetic dataset of 3D point clouds designed as a controlled benchmark for evaluating the stability, robustness, and generalization of loss functions and reconstruction algorithms in point cloud generation and completion tasks. The dataset includes four analytic primitives—sphere, cube, pyramid, and Gaussian cluster—sampled as normalized point clouds and provided under five distinct input regimes simulating sparse, dense, mixed, and incomplete observations. PCOU3D enables reproducible experiments on loss design, reconstruction accuracy, and robustness under varying sampling conditions.

[FIN] Finnish
PCOU3D (PointCloud Oulu 3D) on synteettinen 3D-pistepilviaineisto, joka on suunniteltu kontrolloiduksi vertailuaineistoksi häviöfunktioiden ja rekonstruktioalgoritmien vakauden, luotettavuuden ja yleistettävyyden arviointiin pistepilvien muodostamis- ja täydennystehtävissä. Aineisto sisältää neljä analyyttista perusmuotoa — pallon, kuution, pyramidin ja Gaussin klusterin — jotka on näytteistetty normalisoituina pistepilvinä ja jaettu viiteen eri tiheyteen ja puutteeseen perustuvaan havaintotilanteeseen. PCOU3D mahdollistaa toistettavat kokeet menetelmien tarkkuuden ja robustiuden arvioimiseksi eri pistepilvitiheyksillä.

[ES] Español
PCOU3D (PointCloud Oulu 3D) es un conjunto de datos sintético de nubes de puntos tridimensionales, creado como un banco de pruebas controlado para evaluar la estabilidad, robustez y capacidad de generalización de funciones de pérdida y algoritmos de reconstrucción en tareas de generación y completado de nubes de puntos. El conjunto incluye cuatro formas analíticas —esfera, cubo, pirámide y agrupamiento gaussiano— muestreadas y normalizadas, y se ofrece en cinco regímenes de entrada que simulan observaciones dispersas, densas, mixtas o incompletas. PCOU3D permite experimentos reproducibles sobre la precisión y robustez de modelos de reconstrucción 3D.

[GAL] Galego
PCOU3D (PointCloud Oulu 3D) é un conxunto de datos sintético de nubes de puntos tridimensionais, deseñado como un banco de probas controlado para avaliar a estabilidade, robustez e capacidade de xeneralización de funcións de perda e algoritmos de reconstrución en tarefas de xeración e completado de nubes de puntos. O conxunto inclúe catro formas analíticas —esfera, cubo, pirámide e agrupamento gaussiano— mostreadas e normalizadas, dispoñibles en cinco réximes de entrada que simulan observacións esparsas, densas, mixtas e incompletas. PCOU3D permite experimentos reproducibles sobre precisión e robustez en reconstrucións 3D.

============================================================
OVERVIEW
============================================================

The dataset includes four analytic primitives:
  - Sphere
  - Cube
  - Pyramid
  - Gaussian cluster

Each primitive is represented as a 3D point cloud of uniformly sampled points  within the range [-0.5, 0.5]^3. The dataset is divided into three standard splits:
  - train
  - val
  - test

Every split contains complete shapes ("ground truth") and five input regimes  representing different sampling and density conditions. Inside each split:
  - Files in the root folder are the full ground truth point clouds (1024 points).
  - Subfolders A–E contain the corresponding modified inputs for each regime.

All point clouds are stored as NumPy binary files (.npy) containing arrays of  shape (N, 3) and dtype float32. Coordinates are normalized to fit inside  [-0.5, 0.5] in all axes.

============================================================
 INPUT REGIMES
============================================================

Five regimes are provided to simulate different reconstruction and completion conditions:

  A – Sparse input: random subset of 256 points from the ground truth.
  B – Dense input: full ground truth with 1024 points.
  C – Batch-level mix: half of the samples are sparse (A) and half dense (B).
  D – Random mix: each sample is independently sparse or dense with p=0.5.
  E – Completion: contiguous patch removed; input contains 512–1023 points.

All input regimes share the same filenames as the corresponding ground truth  files, enabling direct pairwise loading for supervised or comparative studies.

============================================================
PURPOSE AND APPLICATIONS
============================================================

The dataset was developed for the paper: Sharifipour, Sasan, Constantino Álvarez Casado, Mohammad Sabokrou, and Miguel Bordallo López. "APML: Adaptive Probabilistic Matching Loss for Robust 3D Point Cloud Reconstruction." arXiv preprint arXiv:2509.08104 (2025).

PCOU3D enables the controlled evaluation of:
  - Stability of loss functions under variable point densities.
  - Sensitivity to partial or missing regions.
  - Consistency across training regimes.
  - Performance under deterministic and probabilistic sampling conditions.

Beyond loss function analysis, the dataset may serve as a benchmark for:
  - Point cloud completion and generation.
  - Shape reconstruction and autoencoder evaluation.
  - Few-shot learning or domain adaptation in synthetic-to-real studies.

============================================================
LICENSE AND USAGE
============================================================

License: CC BY 4.0 (Attribution)
You are free to share and adapt the dataset, provided that proper credit is given 
to the authors and the University of Oulu.

Please cite this dataset as:

  Álvarez Casado, C., Sharifipour, S., & Bordallo López, M.
  (2025). PCOU3D Synthetic Dataset [Data set]. 
  University of Oulu, Center for Machine Vision and Signal Analysis (CMVS).
  DOI: 10.5281/zenodo.17390529

============================================================
ACKNOWLEDGMENT
============================================================

This work was supported by the Business Finland WISEC project (Grant 3630/31/2024), the University of Oulu and the Research Council of Finland (former Academy of Finland) 6G Flagship Programme (Grant Number: $346208$), as part of ongoing research in robust 3D reconstruction  and multimodal sensing.