🚀 Depth Anything V2 - Pro Video Processor

Robust. Hardware-Agnostic. Advanced.

A professional implementation of monocular depth estimation for video, optimized for Google Colab stability and performance.

✨ Features

• Dual Modes: Standard Relative Depth (Visuals) & Metric Depth (Measurements)
• Hardware Smart: Automatically uses GPU (FP16) for speed or CPU (FP32) for compatibility
• 3D Snapshots: Export high-quality 3D Point Clouds (.ply) from any frame
• Robust Engine: Flicker reduction, high-quality FFmpeg encoding, and memory safety
• Flexible Input: Upload videos directly or download from YouTube/URLs

🎯 Quick Start

1. Click the "Open in Colab" badge above
2. Run Cell 1 to install dependencies
3. Run Cell 2 to initialize the depth engine
4. Configure your settings in Cell 3 and run the dashboard

⚙️ Configuration Options

Model Types

Type	Description
Relative	Best for visual depth maps and artistic effects
Metric	Outputs actual depth measurements for 3D export

Model Sizes (Relative Mode)

Size	Description
small	Fastest, lower memory usage
base	Balanced performance
large	Best quality, higher memory usage

Output Resolutions

Option	Description
Native	Original video resolution
720p	HD resolution
480p	Standard definition (recommended for T4 GPUs)
360p	Fastest processing

🧊 3D Point Cloud Export

Generate 3D point clouds (.ply files) from any frame in your video:

1. Set GENERATE_SNAPSHOT = True
2. Set SNAPSHOT_TIME to the desired timestamp (in seconds)
3. The .ply file will be automatically downloaded after processing

Point clouds can be viewed in software like MeshLab, Blender, or CloudCompare.

📂 Project Structure

File	Description
depth_launcher.ipynb	Main Colab notebook with all-in-one setup

📦 Dependencies

The notebook automatically installs all required packages (PyTorch is pre-installed in Colab):

• transformers (from source) - Hugging Face model loading
• accelerate - Optimized inference
• opencv-python - Video processing
• yt-dlp - YouTube/URL video downloading
• pillow - Image processing
• numpy - Numerical computations

💡 Tips

• Use 480p resolution for T4 GPUs to avoid memory issues
• The small model provides good results with faster processing
• Enable temporal smoothing (default: 3 frames) for flicker-free output
• Metric mode is recommended when exporting 3D point clouds

🙏 Acknowledgments

This project uses the Depth Anything V2 model from Hugging Face Transformers.