Skip to content

Gradual selection

Status: stub — populated as part of the alignment-quality wave. Last reviewed: 2026-05-22 (terminology verified against the local Metashape Pro 2.3 user manual PDF).

2.x rename: Metashape 2.x has renamed the menu entry to Clean Tie Points (under Tools → Tie Points → Clean Tie Points…). The criteria-and-threshold dialog itself is unchanged. We keep "Gradual selection" as the page title because that is the term used overwhelmingly in the historical forum content this manual draws on; readers searching for "gradual selection metashape" should land here.

What it is

Gradual Selection (since Metashape 2.x: Clean Tie Points) is the tool used to mark tie points in the tie-point cloud for deletion based on a numerical criterion. It is the primary mechanism for cleaning up an alignment before running Optimize Cameras. The four built-in criteria are Reconstruction Uncertainty, Reprojection Error, Image Count, and Projection Accuracy; this page focuses on the first two, which are the ones most relevant to alignment quality.

Where it lives

Standard Pro
GUI (2.x) Tools → Tie Points → Clean Tie Points… (same)
GUI (1.x) Edit → Gradual Selection… (older) or Model → Gradual Selection… (same)
Python n/a Metashape.Chunk.cleanTiePoints(...) (high-level), or Metashape.TiePoints.Filter().init(chunk, criterion=...).selectPoints(threshold) followed by chunk.tie_points.removeSelectedPoints() for the criterion-driven low-level workflow. Pro only — Python is Pro-exclusive. The 1.x class name Metashape.PointCloud.Filter was renamed to Metashape.TiePoints.Filter in the 2.x API rework.
Available since early PhotoScan (the criteria definitions have been stable since at least 0.9) Python equivalents shifted with the PointCloud / TiePoints API rework in Metashape 2.x — old scripts referencing PhotoScan.PointCloud.Filter will not run on 2.x without rewriting.

Documented in the official manual

The Metashape user manuals describe Gradual Selection in the General workflow → Optimization section, with a list of criteria and short definitions. The forum offers substantially more depth on what the criteria mean geometrically and how to use them in practice — see the threads below.

  • Optimize Cameras — the bundle adjustment step run after gradual selection.
  • Reconstruction Uncertainty and Reprojection Error — the metrics most articles in this manual use.
  • See also: Tie point cloud.

Articles in this manual

Forum threads worth reading

Date Version Author Thread One-line takeaway
2012-10-18 PhotoScan 0.9 Alexey Pasumansky clarification on gradual selection parameters please Reconstruction Uncertainty is a max/min variance ratio in 2-camera triangulation; Reprojection Error is in pixels.
2012-11-15 PhotoScan 0.9 gEEvEE (Geert) clarification on gradual selection parameters please The widely-cited 4-step workflow: align → crop → gradual selection on reprojection (~1 px) → optimize, iterate at most 1–3 times.
2012-11-15 PhotoScan 0.9 Alexey Pasumansky clarification on gradual selection parameters please Never delete so many points that alignment can break.

Caveats

  • The Reprojection Error histogram is in pixels and is therefore comparable across projects. The Reconstruction Uncertainty value is a dimensionless ratio and is mostly useful within a single project, not across projects.
  • The community "1.0 pixel" target is a heuristic, not a universal threshold. Some datasets will not converge below 1.5 px without losing too many points; some converge to 0.4 px easily. Use the iteration count and the chunk's reported average error as the actual feedback.