shriarul5273/FoundationStereo_demo

🔍 FoundationStereo: Zero-Shot Stereo Matching

This is a Gradio demo for FoundationStereo, a state-of-the-art stereo depth estimation model that works on rectified stereo image pairs without requiring dataset-specific training.

🚀 Features

• —Zero-shot stereo matching on rectified image pairs
• —Two model variants available:
• —Low-cost variant (11-33-40): ~750MB, faster inference
• —High-quality variant (23-51-11): ~3.1GB, higher accuracy
• —Automatic model download from Hugging Face Hub
• —Disparity visualization with color-coded depth maps
• —3D point cloud generation with camera calibration
• —Memory-optimized for ZeroGPU infrastructure

📋 Usage

1. 1.Select a model variant from the dropdown (models download automatically)
2. 2.Upload stereo images: Left and right rectified images
3. 3.Process: Get disparity maps and depth visualizations
4. 4.Advanced mode: Generate 3D point clouds with camera parameters

⚠️ Important Notes

• —Rectified images required: Images must be rectified (epipolar lines horizontal)
• —GPU acceleration: Runs on ZeroGPU for fast inference
• —Memory management: Smart caching minimizes GPU memory usage
• —Supported formats: PNG, JPG, and other common image formats

🔗 References

• —FoundationStereo Paper
• —Project Website
• —GitHub Repository
• —Model Repository

📄 License

MIT License - See the original repository for details.