🌲 Granulo-10k

A large-scale benchmark dataset for multiple-view industrial granulometry of OSB wood strands

Granulo-10k is an open benchmark dataset for research on Oriented Strand Board (OSB) strand analysis, with a focus on multiple-view granulometry, strand segmentation, and 3D geometric estimation.

---

🧭 Overview

Granulo-10k contains high-resolution paired images of OSB wood strands acquired with a calibrated two-camera setup, together with segmentation masks, granulometric ground truth, and 3D point clouds.

The dataset supports reproducible research on the estimation of:

• height (h)
• width (w)
• thickness (t)
• compliant / non-compliant strand classification

Granulo-10k addresses a key limitation in industrial wood-vision research: many existing methods rely mainly on 2D projections and do not provide individual strand-level thickness measurements. This dataset provides synchronized multiple views and point-cloud information to encourage research on full 3D strand characterization.

---

✨ Highlights

Feature	Description
🖼️ RGB images	9,600 images at 1280 x 960 resolution
🌲 Strands	200 unique OSB wood strands
🔁 Acquisitions	24 acquisitions per strand
📷 Views	2 synchronized camera views per acquisition
🎯 Masks	Segmentation masks for each paired acquisition
☁️ 3D data	Point clouds associated with paired acquisitions
📏 Ground truth	Height, width, and thickness measurements
🏷️ Labels	Compliant / non-compliant strand categories
🧪 Protocol	Strand-disjoint evaluation to avoid train-test leakage

---

🖼️ Visual Examples

---

🏗️ Acquisition Setup

Images were collected using a calibrated multiple-view acquisition system composed of:

• two synchronized Sony SX90CR color cameras
• a trigger mechanism connected to a photocell
• four LED bars for approximately uniform illumination
• calibrated camera geometry for multiple-view reconstruction

The two cameras were placed at the same height and oriented at approximately 85° with respect to the support, with a camera distance of 125 mm. LED bars were placed at approximately 90 mm from the cameras.

Each strand was dropped from random positions above the cameras while ensuring that it fell inside the intersection of the two fields of view. To increase acquisition variability, each strand was acquired 24 times:

200 strands x 24 acquisitions x 2 cameras = 9,600 RGB images

---

📦 Dataset Content

For each paired acquisition, Granulo-10k provides:

Granulo-10k/
├── Images/        # RGB images from synchronized camera views
├── Masks/         # Segmentation masks for each view
├── PCs/           # 3D point clouds
└── README.md      # Dataset description and citation

Each acquisition includes:

• RGB image from camera 1
• RGB image from camera 2
• segmentation mask for camera 1
• segmentation mask for camera 2
• associated 3D point cloud
• ground-truth height, width, and thickness measurements
• compliance category

---

🌲 Strand Categories

The dataset includes 200 manually measured strands divided into two classes:

Category	Number of strands	Reference average size
✅ Compliant	100	h x w x t = 115 x 20 x 0.70 mm
⚠️ Non-compliant	100	h x w x t = 91 x 9 x 0.65 mm

For each strand, maximum height and width were measured using a caliper. Since thickness can vary across the strand surface, multiple thickness measurements were collected at different points and averaged.

---

📥 Download

The dataset can be downloaded from Hugging Face:

from datasets import load_dataset

dataset = load_dataset("AngeloUNIMI/Granulo-10k")

Dataset page:

https://huggingface.co/datasets/AngeloUNIMI/Granulo-10k

---

🎯 Tasks

Granulo-10k supports several research tasks:

1. Strand segmentation

Estimate binary masks for OSB strands from single-view or paired RGB images.

2. Multiple-view granulometry

Estimate strand height, width, and thickness using one or more available modalities:

• image-only input
• point-cloud-only input
• fused image and point-cloud input

3. Compliant vs. non-compliant classification

Classify strands according to whether their geometry is compliant with manufacturer reference dimensions.

4. Multi-modal geometric learning

Develop methods that combine synchronized RGB views and 3D point clouds for robust geometric reasoning.

---

🧠 Baseline Workflow

The accompanying paper evaluates modern visual backbones and multi-modal fusion strategies for joint estimation of height, width, and thickness.

graph LR
    A[Camera 1 RGB] --> B[Image Encoder]
    C[Camera 2 RGB] --> D[Image Encoder]
    E[3D Point Cloud] --> F[PointNet++ Encoder]
    B --> G[Feature Fusion]
    D --> G
    F --> G
    G --> H[MMoE Decoder]
    H --> I[Height Regression]
    H --> J[Width Regression]
    H --> K[Thickness Regression]

The baseline uses:

• two frozen image encoders, one for each camera view
• a PointNet++ encoder for point-cloud features
• an MLP adapter to align point-cloud and image embeddings
• max-pooling feature fusion
• a Multi-gate Mixture-of-Experts decoder
• task-specific heads for height, width, and thickness regression
• a multi-task uncertainty-weighted loss

---

📊 Representative Results

Values are reported as mean ± standard deviation.

Backbone	Point cloud	Height MAE [mm]	Height MAPE [%]	Width MAE [mm]	Width MAPE [%]	Thickness MAE [mm]	Thickness MAPE [%]
Mean value baseline	-	17.88 ± 1.45	21.58 ± 2.86	6.24 ± 0.50	56.22 ± 3.56	0.18 ± 0.03	29.29 ± 2.32
DINO ViT-B/14	No	2.70 ± 0.34	2.99 ± 0.34	1.65 ± 0.21	12.13 ± 1.67	0.09 ± 0.01	13.70 ± 0.67
ConvNeXtV2	No	2.95 ± 0.31	3.30 ± 0.37	1.64 ± 0.13	11.67 ± 1.16	0.09 ± 0.01	14.91 ± 1.52
EVA02-CLIP ViT-L/14	No	2.82 ± 0.34	3.19 ± 0.35	1.73 ± 0.10	12.11 ± 0.70	0.09 ± 0.00	14.63 ± 1.33
CLIP ViT-L/14	No	2.99 ± 0.24	3.34 ± 0.16	1.84 ± 0.17	13.38 ± 1.69	0.11 ± 0.01	17.77 ± 1.67
DINO ViT-B/14	Yes	2.53 ± 0.08	2.77 ± 0.09	1.59 ± 0.02	11.94 ± 0.35	0.10 ± 0.00	14.58 ± 0.57
ConvNeXtV2	Yes	2.93 ± 0.16	3.24 ± 0.16	1.79 ± 0.06	13.94 ± 0.38	0.10 ± 0.00	15.13 ± 0.13
EVA02-CLIP ViT-L/14	Yes	3.01 ± 0.07	3.31 ± 0.08	1.78 ± 0.07	13.47 ± 0.93	0.11 ± 0.00	17.44 ± 0.79
CLIP ViT-L/14	Yes	3.41 ± 0.14	3.80 ± 0.12	2.06 ± 0.17	15.93 ± 2.02	0.13 ± 0.01	19.96 ± 1.19

DINO ViT-B/14 with point-cloud input achieves the strongest performance for height and width estimation, while thickness estimation remains the most challenging target.

---

🔗 Related Work: IPAN_3D

Granulo-10k is closely related to the earlier work on image-processing-based 3D granulometry.

The related repository provides MATLAB source code for the 2019 IEEE Transactions on Industrial Informatics paper:

3-D granulometry using image processing
R. Donida Labati, A. Genovese, E. Muñoz, V. Piuri, and F. Scotti
IEEE Transactions on Industrial Informatics, vol. 15, no. 3, pp. 1251-1264, March 2019.

Useful links:

• Code: https://github.com/AngeloUNIMI/IPAN_3D
• Project page: http://iebil.di.unimi.it/projects/ipan
• Paper: https://ieeexplore.ieee.org/document/8411142

It can be considered a methodological precursor for multiple-view industrial granulometry, while Granulo-10k provides a larger benchmark dataset for modern learning-based methods using synchronized images, masks, point clouds, and strand-level granulometric ground truth.

---

📚 Citation

If you use Granulo-10k in your research, please cite the associated paper:

@inproceedings{coscia2026granulo10k,
  title     = {Granulo-10k: A Large-Scale Benchmark Dataset for Multiple-View Industrial Granulometry},
  author    = {Coscia, Pasquale and Genovese, Angelo and Piuri, Vincenzo and Scotti, Fabio},
  booktitle = {Proceedings of the IEEE International Conference on Image Processing (ICIP)},
  year      = {2026}
}

If you refer to the related IPAN_3D method or code, please also cite:

@Article{ipan3d,
  author  = {R. {Donida Labati} and A. Genovese and E. Mu\~{n}oz and V. Piuri and F. Scotti},
  title   = {3-D granulometry using image processing},
  journal = {IEEE Transactions on Industrial Informatics},
  volume  = {15},
  number  = {3},
  pages   = {1251--1264},
  month   = {March},
  year    = {2019},
  note    = {1551-3203}
}

---

🙏 Acknowledgements

This work was supported in part by the EC under project EdgeAI (101097300). Project EdgeAI is supported by the Chips Joint Undertaking and its members, including top-up funding by Austria, Belgium, France, Greece, Italy, Latvia, the Netherlands, and Norway.

The authors thank IMAL s.r.l., San Damaso, Modena, Italy, for cooperation in providing data and sample classification. The authors also acknowledge Prof. Ruggero Donida Labati for his contribution to the data collection process.

---

📄 License

This repository is released under the GNU General Public License v3.0.

See the LICENSE file for details.