Monocular Depth Estimation - Inferring Distance from One Image
Monocular depth estimation predicts the distance from a camera to every visible pixel using a single image. Unlike stereo vision, it has no direct geometric triangulation, so the model must learn visual cues such as perspective, texture, object size, shadows, and occlusion.
Relative vs Metric Depth
Relative depth ranks what is closer or farther:
floor near camera < chair < wallIt is useful for image editing and scene layout but does not guarantee meters. Metric depth predicts an absolute value, such as 2.4 m, and needs camera calibration, suitable training data, or additional sensors. Mixing these meanings is a common deployment error.
Learning Approaches
Supervised Learning
Train on images paired with LiDAR, structured-light, or manually produced depth maps. This can achieve accurate metric predictions in a familiar domain, but collecting labels is expensive and sensors have their own gaps.
Self-Supervised Learning
Use adjacent video frames or stereo pairs. A model predicts depth and camera motion, then reconstructs one view from another. Reconstruction error provides the training signal without dense depth labels.
Foundation Models
Large depth models trained across varied imagery can generalize well for relative depth. They are convenient starting points, but their output scale and failure patterns must still be calibrated for the target camera.
Evaluation
Common metrics include:
| Metric | Meaning |
|---|---|
| Abs Rel | Average relative distance error |
| RMSE | Penalizes large absolute errors |
| $\delta < 1.25$ | Fraction of predictions sufficiently close to truth |
Evaluate by scene type, lighting, distance range, and camera configuration. A single aggregate score can conceal dangerous failures on glossy surfaces, thin objects, or low light.
Applications
Depth supports robot navigation, augmented reality, background blur, obstacle awareness, and 3D scene understanding. It is not a substitute for a safety-rated range sensor in applications where an incorrect distance can cause harm. Treat predicted depth as an uncertain estimate, fuse it with other sensors when possible, and define a conservative behavior for low-confidence regions.