Doctoral research · Ongoing
Monocular metric
depth estimation.
Estimate physical depth from a single RGB image. For systems acting in the real world, the question extends beyond producing a convincing map: are the numerical distances dependable when the true depth is unavailable?
Problem setting
One RGB image.
Depth in metres.
No reference at deployment.
Recover physical distance
from one image.
- Input
- A single RGB image of a scene.
- Output
- A depth map in metres, measuring distance along the camera's optical axis. Metric depth assigns physical distances rather than only ordering points from near to far.
- Deployment setting
- Robotic or UAV perception, where the image is available but a reference depth measurement may not be.
A distance that affects
a physical action.
A UAV approaching a landing site needs physical distances to plan its approach. A robot reaching toward an object needs them to plan motion. Errors in numerical scale can change the action, even when the predicted scene looks plausible.
The setting is especially demanding when a system must rely on its estimate without access to the correct depth for comparison.
Visual appearance does not
uniquely specify distance.
Single-image ambiguity. Similar image appearances can arise from different physical scenes, distances, and camera conditions.
Changing operating conditions. Viewpoint, scene content, and imaging conditions may differ from the data on which a system was evaluated.
Limited evidence for an individual prediction. Good average benchmark performance does not establish that one particular estimate is numerically correct. At deployment, the reference answer is often unavailable.
The central question
How dependable is a depth estimate
when there is no reference to compare against?