Depth Estimation
Working out how far away things are from an image — and from a single photo, the absolute scale is mathematically unknowable.
When not to use it
- For absolute measurement, from one image. Scale is destroyed by projection. This isn't fixable.
- For robot navigation or obstacle avoidance, alone. A dollhouse and a room are pixel-identical.
- On reflective or transparent surfaces. The model predicts the reflection's depth, confidently.
- Expecting stereo accuracy at distance. It degrades with distance squared.
Reach for something else instead
- Stereo — a known baseline is an external ruler. Real metres, good to a few of them.
- LiDAR / time-of-flight — measure it. Expensive and correct.
- Structure from motion — many views, and you get scale if you know the camera's movement.
- A known reference object — if there's an A4 sheet in frame, you have scale.
Sources & further reading
- Ranftl et al. (2022), Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer — MiDaS; the scale-invariant loss that encodes the impossibility.
- Godard et al. (2019), Digging Into Self-Supervised Monocular Depth Estimation — monodepth2; supervision from geometry, no labels.
- Eigen, Puhrsch & Fergus (2014), Depth Map Prediction from a Single Image using a Multi-Scale Deep Network — the paper that started it.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reading monocular depth as metres. It's relative, and the absolute is unknowable from one image.
- Building navigation on it. That's the case where the ambiguity has consequences.
- Thinking a bigger model fixes scale. Projection destroyed the information; there's nothing to recover.
- Treating the output as a measurement. It's a prior about what scenes usually look like.
At a glance
FieldComputer Vision
What it givesrelative depth; ordering, not metres
Why absolute is impossibleprojection destroys scale; a room and a dollhouse are pixel-identical
The elegant bitMiDaS's scale-invariant loss encodes the impossibility and unlocks the data
Neverbuild a robot on it alone
DifficultyIntermediate
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Monocular vs. stereo depth — one is a very good prior about what scenes usually look like; the other is a measurement with a ruler. Only one is safe to drive into a wall with.