Media Depth

Context: $ARGUMENTS

Monocular depth estimation — estimate a per-pixel depth map from a single RGB image or video, no stereo rig or LiDAR required. Two actively maintained open-source model families with commercial-safe licenses: Depth-Anything v2 (Apache 2.0, current SOTA, 2024) and MiDaS v3.1 (MIT, Intel ISL, long-standing baseline).

Quick start

• Depth map from a photo (PNG): → Step 3 (depth.py image --model depthanything-v2)
• Depth sequence from a video: → Step 4 (depth.py video)
• 2D photo → side-by-side stereo 3D pair: → Step 5 (depth.py stereo --baseline 6.5)
• Animate a 2.5D parallax / Ken Burns shot from a still: → Step 6 (depth.py parallax --frames 30)
• Pre-download model weights: → Step 2 (depth.py install <model>)

When to use