Skip to main content
Software & Processing

WebODM 3.3.0 and COLMAP 4.2.0: What Shipped, Who Upgrades

WebODM 3.3.0 adds Gaussian splat viewing and automatic datum grid downloads. COLMAP 4.2.0 reworks the global mapper, and 4.2.1 already patches it.

Eric By — M.S. Geography (GIS spec.), FAA Part 107
WebODM 3.3.0 and COLMAP 4.2.0: What Shipped, Who Upgrades

September 2026 brought two open-source releases that matter to drone mappers: WebODM 3.3.0 and COLMAP 4.2.0. WebODM’s biggest change is Gaussian splat support, and COLMAP 4.2.0 was superseded by 4.2.1 four weeks later. This recap follows the August WebODM 3.2.8 and ODM 3.6.2 roundup and is compiled entirely from the GitHub release notes, pull requests, and official docs — nothing here comes from hands-on testing.


WebODM 3.3.0: Splats, Grid Shifts, and a 3D Selector

WebODM 3.3.0 (github.com/WebODM/WebODM, published 2026-09-05 18:17 UTC) lists nine pull requests. Four change what you see or get.

Gaussian splat support (PR #2017). The PR adds four things: downloading training data for Gaussian splats, uploading splat files, automatic tiling of those files into the .RAD format, and displaying splats in the 3D model view where you can “make measurements, use the standard Potree tools.” WebODM does not train the splat. The official docs spell out the loop: process an RGB dataset, download the training .zip from the task’s Splats menu, train it in external software such as OpenSplat or Brush, upload the result. The docs also carry a warning every mapping pilot should read first: “A nadir-only flight, usually yields very poor results.” The Gaussian splatting primer covers why.

Grid shift support and migration (PR #2012, #2016). A grid shift file is a correction grid that PROJ — the coordinate transformation library underneath most GIS software — needs for datum transformations that a simple formula can’t model accurately. If your task’s coordinate reference system needs one and the file isn’t installed, PROJ can’t use its best available transformation and positions shift. Per the PR, WebODM “will now automatically download the appropriate grid files on-demand (if network connectivity is available),” and a database migration goes back through existing tasks, “retroactively downloading the missing grid files and updating the task’s data.” The migration code re-derives each affected task’s raster extents and adjusts its crop polygon. PR #2016 keeps the grids in sync across the webapp and worker containers. The coordinate systems guide explains why the datum step bites.

3D type selector (PR #2005). The 3D view gets “a selector for switching between point clouds and textured 3D models, similar to Map View.”

Bilinear resampling for orthophoto tiles (PR #2013). The tiler now displays orthophoto tiles with bilinear resampling. The maintainer’s own note: “Looks better, at the cost of ~30-40% slower rendering (on my machine at least).”

WebODM 3.3.0 changePRWhat it means
Gaussian splat training-data export, upload, viewing#2017Splats sit alongside point clouds and meshes; training stays external
Grid shift auto-download + task migration#2012Missing PROJ grids fetched on demand; older tasks corrected
Grid sync across containers#2016Workers and webapp share grids; restored on restart
3D type selector, faster Potree menus#2005Switch point cloud / textured model in one view
Bilinear tile resampling#2013Smoother orthophoto display, slower tile rendering
shlex replaces removed pipes module#2002Unblocks Python 3.13 and newer
pdal.bat lookup; subprocess quoting, Electron CSS#2019, #2020Windows and desktop-app fixes

Upgrade notes. Run the upgrade with internet access. The grid migration is best-effort: it logs a warning and skips any task whose grid file it can’t download, so an offline upgrade leaves those tasks’ stored extents and crop polygons uncorrected. No newer WebODM release exists as of October 1, 2026. Full list: the v3.3.0 release notes.


COLMAP 4.2.0: The Global Mapper Stops Throwing Data Away

COLMAP is the structure-from-motion and multi-view stereo package many Gaussian splatting and NeRF pipelines use to solve camera positions. It is not part of the WebODM stack. The 4.2.0 GitHub release was published 2026-09-01 05:44 UTC; the project’s CHANGELOG dates it 08/31/2026.

Multi-component global mapping (PR #4589) is the headline. Before 4.2.0, the global mapper kept only the largest connected group of images and discarded the rest. The release notes: “Disconnected view-graph components are reconstructed independently and returned as separate models,” including components that only separate after outlier rotations are filtered. A site flown as two blocks with no matches between them now yields two models instead of one. The PR author’s benchmark across the 13 IMC2025 scenes reports registered images rising from 634 to 1,544 of 1,823. No drone-survey benchmark accompanies the PR.

This is also a breaking change. The global mapper “may therefore return multiple models for disconnected datasets.” Any script that assumes one output model needs a check. To restore the old behavior, set GlobalMapper.multiple_models to false.

The rest of the release, filtered for mapping relevance:

COLMAP 4.2.0 changeRelease-note detailWho cares
Multi-component global mapper (#4589)On by default; GlobalMapper.min_model_size drops small modelsAnyone scripting global_mapper
Analytical reprojection Jacobians”approximately 1.2—1.55x faster incremental mapping for common pinhole and OpenCV camera configurations”Anyone running incremental mapping with pinhole or OpenCV camera models
ROCm/HIP for patch_match_stereoDense reconstruction on supported AMD GPUs; CUDA and HIP builds are mutually exclusiveAMD GPU owners (HIP_ENABLED build option)
Faster image warping”reported 4—7x improvements in common undistortion cases”Anyone undistorting images
Spatial matching fixFixed for “missing pose priors and large projected coordinates”GPS-tagged drone imagery
Graceful shutdownFirst SIGINT or SIGTERM “preserves usable intermediate results”Long unattended runs
hierarchical_mapper option renameFour options now prefixed HierarchicalMapper.Existing scripts break

Skip 4.2.0 and install 4.2.1. COLMAP 4.2.1, published 2026-09-29, restores “legacy bundle-adjustment behavior in the global mapper, fixing a regression introduced in 4.2.0.” It also fixes a hash-map backend problem that “made same-source builds ABI-incompatible and could corrupt memory,” and a Delaunay meshing loop that never terminated when a 3D point coincided with a camera center. Full list: the 4.2.0 release notes.


Also in the Window: ODX, NodeODM, QGIS

Three smaller releases landed in the same month. Which of the first two applies depends on which side of the ODM/ODX fork you run.

ReleasePublishedNotable
ODX 3.8.32026-09-04WebODM’s engine. build-overviews removed “in favor of cog”; MicaSense RedEdge-P calibration fix; GSD estimated “from bottom 10 percentile instead of average”
NodeODM 3.6.22026-09-24First tagged release since v2.2.3; version now matches the ODM release it wraps; “GPU SIFT works”; Intel image tag renamed from gpu.intel to gpu-intel
QGIS 4.2.32026-09-25Point release, published alongside 3.44.15; the GitHub release carries a commit comparison link and no written notes

If you run the OpenDroneMap-side stack on a GPU, read the NodeODM notes: the :gpu image published in October 2025 “was missing a CUDA library, so --feature-type sift silently ran on the CPU.” ODM itself had no release in the window — 3.6.2 from August remains current.


Bottom line: upgrade to WebODM 3.3.0 if you work in a coordinate system that needs datum grids or want splats next to your point clouds, and do it online. On COLMAP, go straight to 4.2.1 and check any script that expects a single model from the global mapper. If you run NodeODM with a GPU, pull 3.6.2, and if you use the Intel image, switch the tag from gpu.intel to gpu-intel.

Eric

Written by Eric

M.S. Geography (GIS specialization) from St. Cloud State University, FAA Part 107. Pacific Northwest-based; active public-sector Blue UAS operator. Geospatial background covering spatial data, remote sensing, and coordinate systems — applied to drone mapping workflows and deliverables.

About Eric →