SLAMForge Desktop 3.2.0-beta.1

SLAMForge Desktop 3.2.0-beta.1

This is the first SLAMForge desktop beta that reconstructs a dense colored surface instead of only
displaying sparse ORB landmarks. Select or drop a video, provide the matching camera calibration,
and the packaged application performs tracking, loop correction, depth inference, multi-view
filtering, and fusion locally.

Downloads

• SLAMForge-Desktop-3.2.0-beta.1-Windows-x64.zip — portable Windows 10/11 x64 package.
• SLAMForge-Desktop-3.2.0-beta.1-Linux-x86_64.AppImage — portable Linux x86_64 application.
• SHA256SUMS.txt — checksums for both packages.

Windows users must extract the complete ZIP before launching SLAMForge Desktop.exe. The package
is unsigned, so SmartScreen may display an unrecognized-publisher warning.

What is new

• Bundled Depth Anything V2 Small and ONNX Runtime; dense inference needs no Python, network access,
  GPU, or separate model download.
• Learned depth is scale/shift aligned to geometric SLAM landmarks in each selected keyframe.
• Adjacent-view consistency checks and deterministic colored voxel fusion remove unsupported depth.
• The result viewer renders the actual scene colors and dense surfaces.
• Results now contain map.ply (dense), sparse_map.ply (landmarks), trajectory.txt, and
  run.log.
• Weak extreme-scale loops can no longer apply the 20–30x route deformation seen in beta.2.

Expectations and limitations

• Processing has a second offline dense stage and therefore takes longer than beta.2.
• A monocular camera cannot determine absolute metric scale; all geometry remains relative scale.
• Accurate intrinsics and distortion parameters are still required.
• Dense learned depth makes walls and large objects continuous and recognizable, but it is not
  survey-grade geometry and can be unreliable on mirrors, transparent objects, extreme blur, or
  imagery far outside the model's training distribution.
• map.ply is a surface point cloud, not a watertight triangle mesh. Monocular depth and pose drift
  can still produce thick walls, duplicated edges, floating layers, or gaps.
• The downloadable package is CPU-only. A discrete GPU is not required.

Release validation

Two complete runs of the 4,757-frame rectified reference video produced byte-identical trajectories
and sparse maps. Each run exported 4,556 poses with no post-initialization lost frames, 598
keyframes, two rigid loop corrections, and an 886,813-point dense surface cloud. On the 954 poses
with finite ground truth, global Sim(3) alignment gave 1.554 m ATE RMSE. These numbers are a
single-sequence regression check, not a claim of survey accuracy or broad benchmark performance.

SLAMForge is GPL-3.0-only. Depth Anything V2 Small and its ONNX conversion are Apache-2.0; ONNX
Runtime is MIT. The differently licensed Depth Anything Base/Large/Giant weights are not included.