Code and data for segmentation of granular material images using synthetic data and deep learning

This SyntheticSoil project is a Unity-based pipeline designed to procedurally generate and annotate realistic granular soil images from granulometric data. It enables the simulation of soil samples based on real-world particle size distributions (PSDs), their segmentation via Unity Perception, and their conversion to standard annotation formats (COCO, Pascal VOC) for training deep learning models such as Mask R-CNN.

---

## 📁 Repository Structure

```
SyntheticSoilProject/
│
├── 0_AnnotationSyntheticSoil/ # Unity project to generate and annotate soil images (top + bottom views) — version A (original project used in the article)
├── 1_solo2coco/ # Converts Unity SOLO annotations to COCO format using pysolotools
├── 2_polygon_vers_rle/ # Converts polygon annotations (x,y) to COCO RLE format
├── 3_json2xml/ # Converts COCO JSON annotations to Pascal VOC (XML)
├── 4_GrainSegNet/ # GrainSegNet model (training & inference code, checkpoints, evaluation scripts)
├── 5_SyntheticSoil/ # Unity project to generate high-resolution images without annotation — *version B (alternate codebase used in article)
├── 6_Annexes/ # Additional examples, outputs and test files
│ ├── 0_AnnotationSyntheticSoil/ # Example scene's images
│ ├── 1_Example_of_solo_data/ # Example outputs from the annotation Unity project (images, masks, SOLO JSON)
│ ├── 2_Example_of_SyntheticSoil_Images/ Example RGB images and instance masks
│ ├── 3_Test_GrainSegNet/ Example test outputs and evaluation of GrainSegNet
│ ├── Real_granular_material.JPG Example real sample photograph
│ └── Some_granulo_desired.txt Example PSD text file with desired granulometry
├── 7_BuildSyntheticSoil/ Builder version containing the full set of particle models used by the author (pre-processed prefabs)
└── README.md
```