Monocular Depth Estimation

Estimate a relative depth map from a single RGB view.

Input

One image set representing the camera view. A camera calibration data component on the input is required to build a point cloud (intrinsics for unprojection).

Output

A relative depth map per frame. When Export point clouds is enabled and calibration is present, a point cloud is produced as well. The point cloud has the correct pinhole geometry but inherits the relative (non-metric) scale of the depth map — calibration does not convert it to absolute distances in millimeters.

Description

The algorithm computes a relative depth estimate from each image (typical display range [0, 255]). The backend is selected with Depth estimator name when several monocular depth models are available (for example learning-based estimators registered in the Suite).

Export point clouds — when enabled, unprojects the relative depth map using camera intrinsics (and undistortion if needed). Requires camera calibration on the input; output scale remains relative, not metric.

Streaming and live use

For an offline image set, run the algorithm once on the full sequence.

For a live camera stream, use Monocular Stream Depth Estimation instead. It processes each incoming frame continuously and keeps the depth map aligned with the color image for live point-cloud preview in the display.

The same depth backend is used in both cases; the stream variant is intended for real-time preview rather than batch processing of a fixed image set.