Point cloud confidence values: what they mean and how to filter

• Status: unverified
• Applies to: Metashape Pro 2.x — and PhotoScan 1.4+ via the same Point.confidence API; also applies to mesh vertex confidence on built models
• Edition: Pro
• Diátaxis: how-to
• Confidence: high
• Last reviewed: 2026-06-05

Confidence: high. The confidence-value semantics are directly attested by Agisoft support with permalinks (2020, 2022). The Point.confidence API and the Filter by Confidence GUI workflow are introspection-confirmed on Metashape 2.2.

Problem

You've built a dense point cloud. Some regions look noisy or spurious — areas where Metashape's matcher had little confidence in the depth values it produced. You want to:

• Understand what the per-point confidence value actually measures.
• Filter or remove low-confidence points programmatically.
• Export confidence as a per-point attribute for downstream tools (CloudCompare, QGIS, GIS pipelines).

Quick answer. Each point's confidence is the number of depth maps that contributed to it — an integer count, not a probability. Mesh vertices use a similar concept but store a float (averaged depth-map count over a vertex neighbourhood). Higher = more cross-camera agreement = more trustworthy.

Context: what Point.confidence measures

The confidence value on each point in chunk.point_cloud.points is the number of depth maps that contributed to that point's generation. It's an integer count, not a probability or quality score:

"For the dense cloud points the confidence value means the number of the depth maps involved to the point generation process. The value is integer." — Agisoft support, 2022-04-07, Metashape 1.8 (permalink)

A point with confidence=5 was independently observed by 5 depth maps; the matcher fused those depth values into one 3D point. A point with confidence=1 came from a single depth map — no cross-camera verification, more likely to be noise.

For the mesh built from the point cloud, the per-vertex confidence is different:

"The confidence value for the mesh vertices carries the statistical information about the averaged number of depth maps used for the model reconstruction in the neighborhood of the given vertex. The value is float." — Agisoft support, same thread

So:

Object	Confidence type	What it measures
Point.confidence (point cloud)	int	Depth-map count for this exact point
Vertex.confidence (mesh)	float	Local-neighbourhood average depth-map count

Filtering low-confidence points (GUI)

The standard cleanup workflow Agisoft support recommends:

"You can use the confidence filter to show only those points that you would like to remove, then select and delete those points and after that reset filter." — Agisoft support, 2020-02-09, Metashape 1.6 (permalink)

Step-by-step:

1. Tools → Point Cloud → Set Confidence Filter…
2. Set the visible range to the high-confidence end first (e.g., 3 to max). Verify visually that the surviving points look correct.
3. Reset, then set the visible range to the low-confidence end (e.g., 0 to 2).
4. Select all visible (Ctrl+A / Cmd+A) → Delete.
5. Reset Filter (or set range back to full) — the surviving high-confidence points become the active cloud.

Filtering low-confidence points (Python)

The 2.x point-cloud API does not expose per-point iteration or a single "delete by confidence" call. The recommended Python approach uses the same setConfidenceFilter primitive as the GUI, paired with the chunk's region or a polygon as the selection:

Demo verified: ✗ — pending Tier 3 reproduction on a real Metashape install.

import Metashape

chunk = Metashape.app.document.chunk
pc = chunk.point_cloud   # in 1.x this was chunk.dense_cloud

threshold = 3   # remove points with confidence < threshold

# Step 1: limit visibility to the to-be-removed confidence range
pc.setConfidenceFilter(0, threshold - 1)

# Step 2: select the visible points by selecting the chunk's full region
#         (selection respects the active confidence filter)
pc.selectPointsByRegion(chunk.region)

# Step 3: delete the selection
pc.removeSelectedPoints()

# Step 4: restore unfiltered visibility
pc.resetFilters()

print(f"Cleanup at confidence < {threshold} done.")

The pattern: the visibility filter narrows the cloud to the deletion candidates; the region selection grabs them all; the remove call deletes them. Subtleties:

• selectPointsByRegion(chunk.region) selects every point currently visible inside the chunk's bounding box. Combined with the confidence filter, this effectively selects the low-confidence points without needing per-point iteration.
• removeSelectedPoints() honours the selection; the confidence filter does NOT need to be in place during the removal call (only during the selection call).

For spatially-bounded cleanup (e.g., remove low-confidence points only inside a polygon shape), use selectPointsByShapes instead:

Demo verified: ✗ — pending Tier 3 reproduction on a real Metashape install.

boundary = next(s for s in chunk.shapes if s.label == "cleanup_zone")
pc.setConfidenceFilter(0, 2)
pc.selectPointsByShapes([boundary])
pc.removeSelectedPoints()
pc.resetFilters()

If you need confidence-range diagnostics without modifying the cloud, use the filter as a visualisation tool only and skip the selection / remove steps. Combined with point_count per filter range you can build a histogram:

Demo verified: ✗ — pending Tier 3 reproduction on a real Metashape install.

pc.setConfidenceFilter(0, 1)
print(f"conf 0-1: {pc.point_count} points")
pc.setConfidenceFilter(2, 4)
print(f"conf 2-4: {pc.point_count} points")
pc.setConfidenceFilter(5, 255)
print(f"conf 5+: {pc.point_count} points")
pc.resetFilters()

(The exact behaviour of point_count against the active filter is best verified empirically; treat the recipe above as a template to confirm on your install.)

Exporting confidence as a per-point attribute

For LAS / LAZ output, confidence is exported via the extra bytes concept introduced in LAS 1.4:

"Confidence is exported to LAS/LAZ formats via the extra bytes concept introduced in the LAS 1.4 specification." — Agisoft support, 2020-10-16, Metashape 1.6 (permalink)

The Python export:

Demo verified: ✗ — pending Tier 3 reproduction on a real Metashape install.

import Metashape

chunk = Metashape.app.document.chunk

chunk.exportPointCloud(
    "/path/output.las",
    source_data=Metashape.DataSource.PointCloudData,
    save_point_confidence=True,   # write confidence as extra-bytes field
    save_point_intensity=True,    # often paired with confidence in QA workflows
    crs=chunk.crs,
)

save_point_confidence=True is the default, but stating it explicitly makes the script's intent clear. The exported field is readable in CloudCompare (Tools → Other → SF → Set SF as intensity) and in QGIS-with-LAS-plugin.

Recommended threshold values

There is no universally-correct threshold; choice depends on the dataset's overlap pattern. Empirical baselines:

Project type	Suggested low-conf threshold	Rationale
High-overlap aerial (≥80% side, ≥60% forward)	3	Most legitimate points see ≥3 depth maps
Medium-overlap aerial (60-80% side)	2	Many edge points only have 2 observers
Low-overlap aerial / handheld terrestrial	2	Aggressive cleanup risks deleting valid points
Drone close-range / orbit	4-5	High redundancy expected; aggressive cleanup is safe

To tune the threshold interactively before scripted use, run the GUI's Tools → Point Cloud → Set Confidence Filter… and move the slider while watching which areas disappear. Pick the threshold value that hides the noise without removing visible geometry.

The classification connection

The Point Cloud → Classify Points workflow respects the confidence filter. If you've filtered to high-confidence-only before classifying ground points, the classifier ignores the hidden low-confidence noise and produces cleaner classes.

"list(range(128)) — creates the list of ints, that defines the classes of the dense cloud (all classes, actually)." — Agisoft support, 2021-04-29, Metashape 1.7 (permalink)

For batch classification using the confidence filter as a pre-clean step (the classify method lives on the PointCloud in 2.x, not on the Chunk):

Demo verified: ✗ — pending Tier 3 reproduction on a real Metashape install.

chunk.point_cloud.setConfidenceFilter(3, 255)   # show only conf >= 3
chunk.point_cloud.classifyGroundPoints(
    max_angle=10.0,
    max_distance=1.0,
    cell_size=50.0,
    erosion_radius=0.5,
)
chunk.point_cloud.resetFilters()

For the parameters of classifyGroundPoints itself, see Ground classification: the erosion radius parameter.

Caveats

• Confidence depends on processing settings. A point cloud built with quality=Low (4× downscaled depth maps) has fewer per-pixel depth observations than a quality=High cloud, so confidence values are systematically lower. Compare confidences only across clouds built with the same parameters.
• Confidence is computed at build time, not stored from matching. If you re-run Build Point Cloud with new depth maps, the confidence values change. Save the LAS/LAZ export if you need historical confidence for QA records.
• Mesh vertex confidence is float, not int. Post-meshing scripts that assume int will misbehave. The mesh path is separate from the point-cloud path; treat them as different data types even though they share the name "confidence."
• GUI confidence filter only affects display. Unless you delete the hidden points, they remain in the cloud — just invisible. For permanent cleanup, follow the show-low → select-all → delete pattern.
• The 1.x → 2.x rename. chunk.dense_cloud became chunk.point_cloud at the 2.0 transition. Old scripts using dense_cloud.points[i].confidence need the rename to work on 2.x.

See also

• Ground classification: the erosion radius parameter — the classifyGroundPoints workflow that pairs naturally with confidence filtering.
• exportPointCloud defaults: GUI vs Python CRS difference — the broader export-CRS gotcha that applies to confidence export too.
• Mesh and point-cloud editing recipes — for further programmatic point-cloud manipulation.
• main/dense_cloud_max_neighbors tweak (in Undocumented tweaks) — affects how many depth maps contribute to each point, which directly affects confidence values.

References

• Metashape Pro User Manual (2.3), ch. 5 Point cloud → Filter Points by Confidence.
• Point cloud editing with confidence filter tool (Agisoft KB) — Agisoft's walkthrough of Calculate point confidence, Filter by Confidence (which hides, not deletes), and the select-then-delete cleanup.
• Metashape Python API Reference (2.3.1): Chunk.point_cloud, PointCloud.setConfidenceFilter, PointCloud.setSelectionFilter, PointCloud.selectPointsByShapes, PointCloud.selectPointsByRegion, PointCloud.removeSelectedPoints, PointCloud.cropSelectedPoints, PointCloud.resetFilters, PointCloud.classifyGroundPoints, PointCloud.point_count, PointCloud.Point.confidence, Chunk.exportPointCloud.
• Forum thread, Confidence math, 2022 — point vs mesh-vertex confidence semantics; integer vs float distinction.
• Forum thread, Remove Points By Confidence, 2020 — canonical filter-then-delete workflow.
• Forum thread, Export confidence as intensity, 2020 — LAS 1.4 extra-bytes export mechanism.
• Forum thread, Removing low confidence dense cloud points, 2021 — combined confidence filter + classification pattern.