Picsellia CV Pipelines

The official public repository for computer vision pipelines powered by Picsellia CV Engine.

This repository centralizes 17 production-ready pipelines for processing datasets and training computer vision models, all managed through the pxl-pipeline CLI and fully integrated with the Picsellia platform.

---

🚀 Quick Start

Installation

pip install picsellia-cv-engine picsellia-pipelines-cli

Authentication

export PXL_API_TOKEN="your-token"
pxl-pipeline login --organization my-org --env PROD --token $PXL_API_TOKEN

Test a Pipeline

cd pipelines/SAM3_Bbox
pxl-pipeline test SAM3_Bbox --run-config-file run_config.toml

Deploy to Picsellia

pxl-pipeline deploy SAM3_Bbox --organization my-org --bump patch

---

📦 What's Inside

🎯 Pre-Annotation Pipelines

Automatically generate annotations using foundation models:

• SAM-3 Bbox/Polygons - Zero-shot segmentation with natural language prompts
• Grounding DINO - Text-guided object detection
• YOLOv8 Pre-annotation - Fast detection with pretrained weights

🚂 Training Pipelines

Fine-tune models on custom datasets:

• YOLOv8 - Real-time object detection (3 variants)
• RT-DETR - Transformer-based detection
• YOLOv7 Segmentation - Instance segmentation
• SAM-2, CLIP, DINOv2 - Foundation models
• Paddle OCR - Text recognition

🔧 Dataset Processing Pipelines

Transform and prepare datasets:

• Dataset Tiler - Split large images with overlap handling
• Albumentations - Image augmentation
• Bounding Box Cropper - Extract regions
• Diversified Extractor - Sample representative data

---

📚 Documentation

Guide	Description
Processing Pipelines Guide	Pre-annotation and dataset processing
Training Pipelines Guide	Training custom models
Developer Guide	Creating custom pipelines
Documentation Index	Full documentation navigation

Pipeline-Specific Documentation

Each pipeline has its own detailed README:

• pipelines/SAM3_Bbox/README.md - Comprehensive parameter guide
• pipelines/SAM3_polygons/README.md - Polygon segmentation
• pipelines/rt_detr/README.md - RT-DETR training
• And more in each pipeline directory

---

💡 Usage Examples

Pre-annotate a Dataset

# 1. Create configuration
cat > run_config.toml << EOF
[job]
type = "PRE_ANNOTATION"

[auth]
organization_name = "my-org"
env = "PROD"

[parameters]
text_prompt = "car,person,bicycle"
threshold = 0.35

[input.dataset_version]
id = "your-dataset-id"
EOF

# 2. Test locally
pxl-pipeline test SAM3_Bbox --run-config-file run_config.toml

# 3. Deploy
pxl-pipeline deploy SAM3_Bbox --organization my-org --bump patch

Train a Custom Detector

# 1. Create training config
cat > train_config.toml << EOF
[job]
type = "TRAINING"

[auth]
organization_name = "my-org"
env = "PROD"

[parameters]
epochs = 100
batch_size = 16
learning_rate = 0.001

[input.dataset_collection.train]
id = "train-id"

[input.dataset_collection.val]
id = "val-id"

[input.model_version]
id = "pretrained-id"
EOF

# 2. Test
cd pipelines/yolov8/training
pxl-pipeline test training --run-config-file train_config.toml

# 3. Deploy for full training
pxl-pipeline deploy training --organization my-org --bump minor

Process Large Images

# Tile satellite/medical imagery
cd pipelines/dataset_tiler
pxl-pipeline test dataset_tiler --run-config-file tile_config.toml

---

🔧 pxl-pipeline CLI Commands

Command	Description	Example
login	Authenticate with Picsellia	pxl-pipeline login --organization my-org --token $TOKEN
init	Create new pipeline from template	pxl-pipeline init my_pipeline --type processing --template pre_annotation
test	Test pipeline locally	pxl-pipeline test <pipeline> --run-config-file config.toml
deploy	Deploy to Picsellia cloud	pxl-pipeline deploy <pipeline> --organization my-org --bump patch

Learn more: Picsellia CV Engine - CLI Usage

---

🏗️ Repository Structure

picsellia-cv-pipelines/
├── pipelines/              # 17 production-ready pipelines
│   ├── SAM3_Bbox/
│   ├── yolov8/
│   ├── rt_detr/
│   └── ...
├── docs/                   # Comprehensive guides
│   ├── PROCESSING_PIPELINES_GUIDE.md
│   ├── TRAINING_PIPELINES_GUIDE.md
│   └── DEVELOPER_GUIDE.md
├── tests/                  # Test configurations
├── scripts/                # Automation scripts
└── .github/workflows/      # CI/CD

Each pipeline includes:

• pipeline.py - Main entry point
• steps.py - Step implementations
• utils/parameters.py - Parameter definitions
• config.toml - Metadata
• Dockerfile - Container definition
• README.md - Detailed documentation

---

🎓 Getting Started Path

For Users:

1. Install CLI tools → See Quick Start
2. Choose your use case:
   • Processing data? → Processing Guide
   • Training models? → Training Guide
3. Check pipeline-specific README for detailed parameters

For Developers:

1. Clone repository: git clone https://github.com/picselliahq/picsellia-cv-pipelines.git
2. Setup: uv sync && pxl-pipeline login
3. Read: Developer Guide
4. Create pipeline: pxl-pipeline init my_pipeline --type processing

---

🛠️ Development

Quick Development Setup

# Clone and setup
git clone https://github.com/picselliahq/picsellia-cv-pipelines.git
cd picsellia-cv-pipelines
uv sync
pre-commit install

# Authenticate
pxl-pipeline login --organization test-account --env STAGING --token $TOKEN

Create a Custom Pipeline

# Initialize from template
pxl-pipeline init my_pipeline --type processing --template pre_annotation

# Customize
cd pipelines/my_pipeline
# Edit utils/parameters.py, steps.py, pipeline.py

# Test
pxl-pipeline test my_pipeline --run-config-file test_config.toml

# Deploy
pxl-pipeline deploy my_pipeline --organization my-org --bump patch

Full guide: Developer Guide

---

🔬 Testing & CI/CD

Run Tests

# Test specific pipeline
./scripts/test_pipelines.sh --pipeline SAM3_Bbox

# Test all pipelines
./scripts/test_pipelines.sh --pipeline all

CI/CD

• Automated testing on every PR
• GPU tests on self-hosted runners
• Code quality checks with Ruff

---

💻 Hardware Requirements

Use Case	Minimum	Recommended
Processing (CPU)	4 cores, 8GB RAM	8 cores, 16GB RAM
Processing (GPU)	8GB VRAM	16GB VRAM (10-20x faster)
Training	8GB VRAM, 16GB RAM	24GB VRAM, 32GB RAM, NVMe SSD

---

🆘 Troubleshooting

Common Issues

Authentication errors?

pxl-pipeline login --organization my-org --token $NEW_TOKEN

Out of memory?

# Reduce batch size in config
[parameters]
batch_size = 4

Slow processing?

# Verify GPU is detected
python -c "import torch; print(torch.cuda.is_available())"

More help: Troubleshooting sections in guides

---

🤝 Contributing

We welcome contributions! Please:

1. Fork the repository
2. Create a feature branch
3. Make your changes (follow existing patterns)
4. Add tests and documentation
5. Submit a pull request

Contribution types:

• New pipelines
• Bug fixes
• Documentation improvements
• Performance optimizations

See: Contributing section in docs

---

📝 License

Apache License 2.0 - See LICENSE

Individual pipelines may use models with different licenses. Check model cards on Hugging Face and framework licenses.

---

🙏 Acknowledgments

Built with:

• Picsellia CV Engine - Core framework
• Meta AI - SAM-3, SAM-2, DINOv2
• Hugging Face - Model hosting and transformers
• Ultralytics - YOLOv8, YOLOv7
• OpenAI - CLIP
• PaddlePaddle - Paddle OCR

---

🔗 Links

• Documentation: Full Documentation Index
• CV Engine Docs: picselliahq.github.io/picsellia-cv-engine
• Picsellia Platform: picsellia.com
• GitHub Issues: Report bugs or request features
• Support: support@picsellia.com

---

Built by Picsellia | Website | Documentation | Blog