SatVision

An opinionated framework for deploying computer vision models on remote sensing imagery.

🚀 Getting Started

The framework supports both local Python inference and scalable deployment via NVIDIA Triton Inference Server.

1. 💻 Local Inference

Best for development and direct integration into Python applications.

Installation:

pip install SatVision
# For TensorRT support (requires NVIDIA GPU and drivers):
pip install "SatVision[tensorrt-cu12]"

Usage Example:

import satvis
from pathlib import Path

# Initialize model (supports 'torch' or 'tensorrt' backends)
# Note: 'tensorrt' backend requires the optional dependency installed above.
model = satvis.get_inference_model(
    model_name="torchvision::resnet50", 
    # override_args={"backend_type": "tensorrt"} # Uncomment to use TensorRT
)

image_path = Path("path/to/your/image.jpg") # Replace with your image path

predictions = model.predict(
    image=image_path, 
    apply_transform=True
)

best_prediction = max(predictions.items(), key=lambda item: item[1])

print(f"Best prediction: {best_prediction}")

2. 🏗️ Server Deployment

Deploy models using NVIDIA Triton Inference Server for high-performance serving.

Setup:

1. Clone the repository:

   git clone https://github.com/liopeer/SatVis.git
   cd SatVis
   

2. Install just (Command Runner): Follow the instructions at Just Docs to install just for running predefined commands.

3. Install Dependencies: This will install uv (package manager) and project dependencies.

   just install-uv
   just install-dev
   source .venv/bin/activate
   

   This installs tensorrt==10.9.0.34. This version is compatible with Nvidia Triton Server nvcr.io/nvidia/tritonserver:25.03-py3 and driver >=570.

4. Generate Models: Create the optimized ONNX/TensorRT models for Triton Server.

   python server/generate_models.py
   

5. Launch Triton Server: Requirements: NVIDIA Driver >= 570, Docker with CUDA Toolkit.

   docker compose -f server/docker-compose.yml up -d
   

6. Run Inference Client: Send an inference request to the running server.

   uv run scripts/predict_resnet_http.py
   

🗺️ Roadmap

Q4 2025

• Core PyTorch Framework with Inference and ONNX Export:
  • Classification
  • Image Embeddings
  • Language Embeddings
• TensorRT Export
  • FP32/FP16 export
  • verification
• Model Serving
  • Nvidia Triton server with kserver API
  • Model Zoo for classification and embeddings
• Documentation of REST API
• Basic CI/CD
  • Inference server to container registry
  • Unit tests with pytest and coverage

Q1 2026

• Model Training
  • Panoptic Segmentation
  • CLIP-style training for image and language embeddings
• Multi-Spectral / Hyperspectral support
  • I/O for various formats (GeoTIFF, HDF5, NetCDF)
  • Data Augmentation for multi-spectral data
  • Models
• Documentation for Python API

Q2 2026

• Model Quantization
  • Int8 PTQ (Post Training Quantization) with calibration

Small Projects

• PostGIS sampler for PyTorch DataLoader

👤 Maintainers

• Lionel Peer (lionel.peer@gmail.com)