Absolute Scale Estimation from Ground Plane and Camera Extrinsics

This direction is merged into another direction: monocular scale recovery. Its last version stays readable and citable here.

Version 1 of 178 papers102 conceptsfrozen 2026-10-04

A single camera is inherently "scale-blind." Because a 2D image is a projection of a 3D world, a small object close to the lens can appear identical to a massive object far away. This phenomenon, known as scale ambiguity, means that traditional monocular systems can reconstruct the shape of a scene but cannot determine if they are looking at a dollhouse or a warehouse.

To bridge the gap between relative pixels and real-world meters, researchers have developed a suite of techniques that "anchor" visual data to physical constants. By leveraging geometric priors, known object dimensions, and inertial sensors, monocular systems can now achieve metric accuracy comparable to expensive LiDAR or stereo-camera setups.

The Ground Plane as a Physical Ruler

The most robust way to recover scale is to treat the ground as a geometric reference. If a camera is mounted at a known height and angle, every pixel on the road surface corresponds to a specific physical distance.

Chunyan Han and Li-Chen Lin demonstrated this in Automobile Distance Measurement System on Highway Based on Vehicle-mounted Camera by using standardized road markings to establish a pixel-to-meter scale factor. Similarly, Robert McCraith and colleagues proposed a vision-only calibration that fits a 3D plane to the ground surface. By comparing this fitted plane to the known mounting height of the camera, the system derives a per-frame scaling factor to convert relative depth into metric units.

This "ground constraint" is now a staple of self-supervised learning. In GroCo: Ground Constraint for Metric Self-supervised Monocular Depth, Aurélien Cecille and Rémi Agier show that enforcing consistency between predicted depth and a theoretical ground plane allows models to learn absolute metric units without ever seeing a LiDAR label.

Known Objects as Semantic Anchors

When the ground is not visible or reliable, systems can look for "semantic anchors"—objects with predictable physical sizes.

  • Vehicles and Pedestrians: Pradipta Ghosh and the team behind CamLoc use detected vehicles as reference markers. Since most cars fall within a standard size range, their pixel dimensions provide a reliable hint for determining camera elevation and distance.
  • Infrastructure: Young Ki Baik and co-authors utilize lane-width priors to resolve camera height and roll, while Peisong Li uses the fixed distance between power lines to monitor tree intrusions with sub-10% error.
  • Wearables: In indoor settings, Jie Hao and Jun Luo 0001 leverage the fixed height of a head-mounted display to derive 3D room dimensions from a single image, bypassing the need for complex multi-view geometry.

Fusing Motion and Inertia

Scale can also be recovered through the physics of motion. While a camera only sees "optical flow" (pixels moving), an Inertial Measurement Unit (IMU) measures actual physical forces like acceleration ($m/s^2$).

Nathaniel Merrill and Marc Pollefeys introduced CodeVIO, which tightly couples learned depth with inertial measurements. By representing depth as a compact "code" within a Kalman Filter, the system uses real-world motion to continuously refine the scale of its 3D maps. Sen Zhang and Jing Zhang took this further with DynaDepth, using a differentiable EKF to fuse visual and inertial data, allowing the model to predict real-world distances directly from motion dynamics.

Learned Metric Priors and Foundation Models

Recent breakthroughs have moved toward "zero-shot" metric estimation, where neural networks are trained on massive, diverse datasets to "know" the scale of the world implicitly.

Metric3D v2, developed by Chi Zhang and Wei Yin, resolves focal length ambiguity to provide accurate metric depth across "in-the-wild" images. These foundation models are now being integrated into SLAM (Simultaneous Localization and Mapping) pipelines. For instance, Jian Yang and Jin Xie use these priors in GigaSLAM to maintain consistent real-world dimensions over kilometer-scale trajectories, preventing the "scale drift" that typically plagues long-range monocular navigation.

By combining these geometric, semantic, and inertial anchors, the research community is transforming the monocular camera from a simple 2D recorder into a precise 3D metric sensor.

Papers in this version

Showing 60 of 78 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_absolute_scale_estimation_from_ground_pl_1f503f05,
  title = {Absolute Scale Estimation from Ground Plane and Camera Extrinsics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/absolute-scale-estimation-from-ground-plane-and-camera-extrinsics/rdr_1f503f05cb224f47b2ca725b33dfcc7a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_absolute_scale_estimation_from_ground_pl_1f503f05_v1,
  title = {Absolute Scale Estimation from Ground Plane and Camera Extrinsics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/absolute-scale-estimation-from-ground-plane-and-camera-extrinsics/rdr_1f503f05cb224f47b2ca725b33dfcc7a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}