Depth Estimation

A photograph is flat, but the world that produced it has distance: a mug sits in front of a laptop, a pedestrian is nearer than the building behind them. Depth estimation is the process of predicting that distance structure from visual data, giving a vision system a sense of which surfaces are close, far, or separated in 3D space.

What the model produces

The result is commonly a depth map: an image-sized grid in which every pixel receives a depth value. Larger or smaller values indicate distance from the camera, depending on the chosen convention. A related representation, disparity, measures how far a pixel shifts between two side-by-side camera views; closer objects shift more. Depth can be estimated from:

Stereo images, by matching corresponding pixels in left and right views and using camera geometry.
Video, by tracking how scene points move as the camera moves.
A single image, using a trained neural network that learns visual depth cues such as perspective, occlusion, shadows, texture size, and familiar object sizes.
Depth sensors, such as LiDAR or structured-light cameras, which directly measure distance and can provide training labels.

Why depth is difficult

A single image does not uniquely reveal real-world scale: a small nearby car can look like a large distant car. Monocular models resolve this ambiguity from learned patterns, so their estimates can be relative rather than exact metric distances. Reflective windows, blank walls, thin objects, and object boundaries also create trouble because they offer weak or misleading visual clues. Stereo methods improve physical accuracy, but require calibrated cameras and reliable matching.

What it enables

Depth estimation helps an autonomous vehicle judge free space and avoid obstacles, lets a robot decide how far to reach, and supports realistic background blur or augmented-reality objects that appear behind a real table rather than floating through it. In medical imaging, depth-like 3D structure helps separate anatomical surfaces; in inspection, it exposes dents or height differences that ordinary color images miss. Practical models include MiDaS for monocular relative depth and RAFT-Stereo for stereo matching; OpenCV also provides stereo tools such as StereoSGBM.

Depth estimation is the task of predicting the distance from a camera to visible scene points, producing a per-pixel depth map or relative depth ordering from one or more images. It can use stereo pairs, video motion, active sensors, or learned monocular models. Depth estimation enables 3D scene reconstruction, object-scale reasoning, robotic navigation, augmented reality, and reliable spatial interaction; without it, image-based systems lack explicit geometric understanding.

Imagine looking out a window and instantly knowing that a nearby tree is closer than the house across the street, while mountains are far away. Depth estimation gives computers a similar sense of distance from what they see.

Instead of treating a photo as a flat picture, it helps an AI estimate how far away each part of the scene is. A self-driving car can use this to judge whether a cyclist is close; a robot can use it to avoid a table; and a phone can use it to blur the background behind a person. It matters because understanding distance helps machines move and react safely in the real, three-dimensional world.