← AI Terminology
Depth Estimation
Depth estimation is the computer vision task of predicting the distance from a camera to each pixel in an image — producing a depth map that encodes the 3D structure of a scene from 2D image input.
It is foundational for autonomous vehicles, augmented reality, robotics, and 3D reconstruction.
It is foundational for autonomous vehicles, augmented reality, robotics, and 3D reconstruction.
Why It Matters in AI
Cameras produce 2D images, but the world is 3D. Depth estimation recovers that missing dimension — enabling robots to navigate around objects, AR to overlay content correctly on surfaces, and self-driving cars to estimate distances to pedestrians. Modern monocular depth models (Depth Anything V2, ZoeDepth) can estimate depth from a single image without stereo cameras or LiDAR.
Key Points
| Aspect | Description |
|---|---|
| Stereo | Depth from two images (disparity) — geometrically grounded, less ambiguous than monocular |
| Metrics | AbsRel (absolute relative error), RMSE, δ<1.25 threshold accuracy |
| Monocular | Depth from a single image — underdetermined problem, requires learned priors |
| Supervised | Train on LiDAR or structured-light ground truth (KITTI, NYUv2 datasets) |
| Self-supervised | Train on video — use view synthesis consistency as training signal (no depth labels needed) |
| Foundation models | Depth Anything V2, ZoeDepth, Marigold — zero-shot depth estimation on arbitrary images |
Simple Analogy
Looking at a photo of a street scene, you intuitively know the car in the foreground is closer than the building in the background — based on size, texture, and overlap cues. Monocular depth estimation trains a neural network to exploit these same monocular cues learned from thousands of images with known depth.
Common Usage Examples
depth_anything_v2.infer_image(image)— zero-shot depth estimation via Depth Anything V2- iOS Measure app and Face ID: depth estimation using structured light on Apple devices
- Tesla Autopilot: camera-based depth estimation instead of LiDAR for 3D scene understanding
AdaBinsmodel: depth estimation on NYUv2 indoor dataset, SOTA with adaptive bin widths- NeRF (Neural Radiance Fields): implicitly learns a 3D scene representation including depth
Summary
In short: Depth estimation converts 2D images into 3D distance maps — enabling machines to understand spatial structure from a single camera, a capability essential for autonomous systems.