Depth Estimation
Inferring how far objects are from the camera - recovering 3D structure from a 2D image.
Think of it like
Humans use binocular vision and visual cues (perspective, shadows, blur) to judge distance. Depth estimation teaches machines to do the same from one image.
Example
Given a single photo of a street, predict how far each building, car, and pedestrian is from the camera. Used in AR apps to place virtual objects correctly on real surfaces.
How it actually works
Monocular depth (single image) is underconstrained - many 3D configurations project to the same 2D image. Models learn statistical priors from training: familiar objects are certain sizes, surfaces follow typical slopes. Stereo depth uses two cameras and triangulation, which is more reliable but requires paired hardware. Combines photometric cues (shading, perspective) with learned priors. Photogrammetry and NeRF approaches also recover depth.
For product teams
Enables AR/VR, robotics, autonomous vehicles, and 3D video editing.
For engineers
Monocular: usually supervised with depth ground truth from LiDAR or structured light. Metrics: absolute relative error, RMSE. Benchmarks: NYU Depth v2, KITTI.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome