depth estimation
/ depth es-tih-MAY-shun /
Depth estimation works out how far away each part of a scene is — turning a flat photo into a sense of near and far. The output is usually a depth map: an image where each pixel's brightness encodes distance, so the foreground glows bright and the background fades dark (or vice versa). It is the difference between seeing a picture of a hallway and understanding that the hallway recedes thirty feet into the distance. This 3D sense is what robots, cars, and AR apps need to move through and interact with the world.
There are two broad ways to get it. The reliable way uses geometry and extra sensors: two cameras spaced apart (stereo, like your two eyes), or a sensor that times how long light takes to bounce back (LiDAR, time-of-flight). The harder, trendier way is monocular depth estimation — guessing depth from a single ordinary photo using a neural network that has learned the cues humans also use: things higher in the frame are often farther, textures shrink with distance, nearer objects block farther ones.
The honest line runs right between those two. Geometric, multi-sensor depth gives real measurements you can trust for safety-critical tasks. Single-image depth is, by contrast, fundamentally a guess — a single flat photo genuinely does not contain enough information to fix true distance (a real mountain and a tiny photograph of a mountain can project identically), and the model only knows relative depth and typical scenes, not absolute meters. It is impressive and useful for effects like portrait-mode background blur, but it is not a measurement, and self-driving systems lean on LiDAR and stereo precisely because they cannot bet lives on a monocular guess.
Your phone's portrait mode estimates depth to find which pixels are the subject and which are background, then blurs the background for a creamy bokeh. It usually works beautifully — but watch it smear the blur across a stray hair or the gap between an arm and the body, where its depth guess gets the boundary wrong.
Single-image depth powers portrait blur — impressive, but a guess, not a measurement.
A single 2D photo cannot, in principle, pin down true distances — depth from one image is an educated guess, typically only relative (nearer/farther), not absolute meters. Safety-critical systems use stereo cameras or LiDAR for real measurement and treat monocular depth as a hint, not ground truth.