Version: 3.4 (unreleased)

Release Notes 3.03

3.3.41 – Photogrammetry Update

July 2026

✨ New & Enhanced Features

Ground Control Points

Ground control points are now first-class project objects: visible and pickable in the 3D view, listed in the Data Tree under Ground Control Points, and saved with the project. They are reusable outside SfM.
Import Ground Control Points (CSV)… reads what survey instruments actually export — byte-order marks, quoted fields, comma/semicolon/tab/space delimiters (auto-detected), and decimal commas. Coordinates can be supplied in the project CRS or as WGS84 longitude/latitude, and are reprojected on import. Any row it cannot read is reported with its line number rather than silently dropped.
Mark control in the photo browser (right-click → Mark Ground Control Point). Three or more marked points then drive the georeferencing in place of camera GPS — centimetre survey instead of metre-level GPS — with GPS demoted to a soft constraint. The log always states which data placed the model.
Points can be flagged as check points: held back from the fit and only measured, which gives an independent estimate of the model's true accuracy.
Seed Mesh Tie Points from Ground Control turns measured control into a mesh's own tie points, so the existing Corrections → GCP tool can register a mesh against a survey.

Scale bars

Scale Bars (Known Distances)… gives a model real size when there is no GPS and no ground control: pick the two tie points either end of something you measured and type the distance. Several bars are combined by median, so one mistyped distance cannot drag the model's scale.
Where GPS or control already set the scale, scale bars are measured as an independent check of it rather than applied.

Quality report

Show Quality Report summarises a reconstruction in one place: registration, reprojection RMS with median/90th/99th percentiles, what georeferenced the model and how well, per-GCP accuracy with control and check points reported separately, scale-bar agreement, and the worst photographs by residual. Written to the messages panel and to a text file in the workflow folder.

Texturing

Textures are now blended from multiple photographs per texel, weighted by how well each one sees that point, which dissolves the exposure and white-balance seams that appeared where neighbouring triangles drew from different photos.
The texture budget is now calculated for you. Texture Mesh from Photos leads with a recommendation worked out from your own data — the surface area of the mesh against what the photographs can actually resolve at their range — so accepting it takes one click and the manual options become an advanced path. Asking for far more texture than the imagery contains is the commonest way to spend gigabytes and hours storing blur, and the prompt now states the resolving power alongside the page count.
Page size is offered as 4096, 8192 or 16384 and clamped to 16384, which is the largest 2D texture AMD and Intel GPUs will create. The page-count prompt is pre-filled from the mesh rather than defaulting to 2.
Two new per-triangle attributes are written by every bake: Texture GSD (units/texel) and Texture resolution 0-1. Colour the mesh by either to see how well resolved each part is; the second is normalised for meshes with no real-world scale. The log reports whether the model is limited by photograph resolution or by texture budget — which is the difference between "fly lower" and "add pages".
Fewer flat facets: the occlusion test's tolerance now scales with triangle size, so a ray clipping nearby micro-relief is no longer mistaken for a real obstruction, and a triangle that loses every candidate is textured from its best photograph rather than left flat. Only triangles no photograph framed at all fall back to vertex colour.

Camera calibration

VRGS now reads the camera make and model from EXIF, and recognises bodies that remove lens distortion in firmware — every DJI drone, and most phones. For these the distortion coefficients are held at zero through bundle adjustment instead of being solved for.
This fixes a subtle and damaging failure. Left free on an already-corrected camera, bundle adjustment still finds distortion coefficients: on a near-nadir survey it can lower reprojection error by absorbing feature-matching noise into the radial terms while the poses compensate. The solution reports an excellent reprojection error and is geometrically wrong, and the error only surfaces when photographs are projected onto real geometry — as smeared, evenly spaced bands across the baked texture.
Every bake now measures whether the distortion model belongs to the imagery, by checking whether overlapping photographs agree when sampled at the same point in space, and warns if the setting is the worse of the two. A correct camera model makes the two readings identical.
When a photograph carries no EXIF focal length, the solve falls back to a 50 mm assumption — roughly 2.5× too long for typical drone or phone imagery, which is enough to stop photographs registering. This used to happen in silence, and looked like a matching problem rather than a missing tag; the run now reports how many photographs were affected and why it matters.

Dense MVS diagnostics

New log lines explain why a dense cloud came out sparse, rather than leaving the fusion funnel as the only clue. A confidence profile shows how much surface sits just below the confidence cutoff, so you can tell whether lowering it would recover real geometry or merely admit noise. A per-reference yield line reports which views produced almost nothing, separating a localised hole from uniform thinning and pointing at the stretch of flight path responsible.
The per-reference line now tags each weak view with its partner counts — how many overlapping views were assigned to it versus how many actually confirmed its depth. This tells a coverage gap (nobody overlaps that viewpoint) apart from a bad camera pose (overlapping views were found but disagree), which need completely different fixes and were previously indistinguishable from a bare point count.
Each view now reports whether its depth search was range-limited (the surface lay outside the depth range searched), texture-limited (nothing to correlate), over-quantised (measured, but sampled too coarsely to confirm at range) or terraced (depths quantised onto shells). The first two are not fixable from the settings panel, so this is worth reading before spending a run adjusting the dense parameters — the guide's Reading the run log section covers how to act on each.
The guide also now explains why the dense cloud has gaps where the sparse cloud has points, which is expected behaviour rather than a fault: dense matching needs three views and a whole patch to correlate, where sparse needs two views and a single feature.

Photogrammetry workflow

Cameras with EXIF GPS now appear in the 3D view as soon as photos are imported, before any reconstruction — drawn without thumbnails to mark them as provisional GPS positions rather than solved poses.
Quality presets (Fast / Balanced / Accurate) over the reconstruction parameters, plus eight previously hidden settings exposed in the Properties panel: match strategy, retrieval threshold and neighbours, epipolar RANSAC, registration tolerance, and both triangulation-angle gates.
The sparse cloud now carries Observations and Reprojection error attributes, so weak camera coverage and bad points can be found by colour and filter — before a dense MVS run spends hours on them.
Open Photos in Annotation View opens a mesh's photographs for SAM2/UNet interpretation, with clicks propagating between photos through the mesh.
Add Satellite Imagery Basemap and Show Camera GPS as Point Cloud added to the workflow menu.
Progress bars for reconstruction and for photo import.
A GPS altitude is MSL option, for cameras that write true MSL rather than the ellipsoidal heights most drones record.

Virtual Field Trips

All the trip controls are in one place. The panel's icon toolbar and the ribbon's All Pointers button are gone; everything is now a labelled button in a single row at the top of the VFT panel, all the same size, wrapping onto more lines as the panel narrows. Each button appears only when it applies to you and lights up when it is on, so there are no controls that do nothing — and no trip controls sitting on the ribbon when no trip is running.
🎥 Follow the leader now works. Camera-follow only ever drove the old OpenGL view, so the button did nothing in the Vulkan view. Your camera now tracks the leader's as they move, in both.
🧽 Clear Ink. The leader clears the board for the whole trip; anyone else clears their own view and withdraws only their own strokes, so no one can erase someone else's work.
🔦 Pointers switches between seeing only the leader's pointer and everyone's, each labelled with its owner.
The trip set-up dialog opens on Join, with the cursor in the invite-code box — being handed a code is the common case. Codes now tolerate the way they actually arrive: any case, and stray spaces, line breaks or separators from an email or chat message. A code that could not possibly be one is refused with a message instead of a three-second wait on a trip that does not exist. The mislabelled Port field (which set a development token, not a port) and a Copy button that copied the trip name have been removed.

🐛 Fixes & Reliability

Dense clouds no longer have holes where the sparse cloud has data. The depth range each view searched was padded in a way that collapsed to a near-zero start whenever the view spanned more than about an 11:1 range front-to-back — common on any oblique or valley shot. The entire depth budget was then spent in the first few centimetres in front of the camera, and the view contributed nothing to the dense cloud. On one 252-image survey this silently emptied a third of the views; fixing it recovered the missing surface and also cut the scattered-debris count roughly twelvefold. The learned (CasMVSNet) backend carried the same fault and is fixed with it. The plane-sweep depth search was also made finer, which the same fix had been masking.
Reconstructions are now georeferenced. The model was previously left in a local frame relative to the photo centroid and never converted to the project coordinate system, so sparse and dense clouds could sit hundreds of kilometres from the camera GPS. Reconstructions are now fitted directly to project-CRS coordinates.
Sparse and dense clouds are saved. They were written to the project database but their point data never reached disk, so they reloaded empty.
Baked textures now appear immediately in the Vulkan view. Previously a re-baked mesh kept the previous unwrap's texture coordinates until the project was reloaded, showing scrambled texture.
Importing a photograph with no GPS data no longer places a phantom camera on the equator once a project coordinate system is set.
VFT ink no longer draws lines shooting off the outcrop into the sky. The vertex layout for plain 3D lines gained a dash attribute and grew by four bytes, but the packer feeding it was never updated, so the graphics card read every vertex slightly further into the buffer than the last and the positions drifted. Ink strokes are now drawn as rounded 3D lines of constant on-screen thickness, and a stroke that slips off the surface it started on breaks cleanly instead of stretching across the scene. The same packing fault also affected the seismic bounding box and the DFN line styles (slicer lines, fracture outlines, intersection lines), which are fixed with it.
Avatars and pointer beams no longer sit on your camera. While following the leader — or just after a recall — your viewpoint is theirs, so their avatar was drawn at your eye and their pointer beam ran straight down your line of sight. Both are now cleared out of the way; the marker at the end of the pointer, which is what tells you where to look, stays.
Invite codes copied from the trip dialog are no longer truncated to a single character by the clipboard.
The Coordinate System dialog now pre-selects the exact EPSG code for the zone covering your photos, instead of filtering by name (which matched every datum's version of that zone).
Ground control spread along a single line is now refused with an explanation, rather than producing a model rolled arbitrarily about that line while looking correctly georeferenced.
Photographs, camera poses and calibrations are attached to the mesh a workflow produces, so they survive save and reload and can be used for interpretation.

🔄 Minor Updates

The GPS pair max distance default is documented correctly as 100 m (it was previously documented as disabled). Photo pairs further apart than this are skipped during matching; raise it, or set 0, if genuinely overlapping long-range shots are failing to register.
Removed a dead "Create Image Tie Points" mode in the photograph view that only left the mouse in a state where clicking did nothing.

3.3.23 to 3.3.40 – Major Update

April – July 2026