Skip to content

CLI binaries

60 ARM64 native aliceVision_* binaries ship with the release tarball / Homebrew formula, covering every node referenced by the upstream 2026.1.0 Meshroom template set. All preserve the upstream AliceVision CLI surface unchanged (no renamed flags). For per-flag detail use -h on the binary itself; this page summarizes what each one does and how it fits the pipeline.

The classic 11-binary photogrammetry chain below is the original Phase 1 scope. The other 49 binaries (modern SfM, HDR, panorama, photometric stereo, LIDAR, camera tracking, utilities, Mac-port-native ML wrappers) were added in Phases 14.1–14.9.

Pipeline order — classic photogrammetry chain

flowchart LR
    A[cameraInit] --> B[featureExtraction]
    B --> C[imageMatching]
    C --> D[featureMatching]
    D --> E[incrementalSfM]
    E --> F[prepareDenseScene]
    F --> G[depthMapEstimation]
    G --> H[depthMapFiltering]
    H --> I[meshing]
    I --> J[meshFiltering]
    J --> K[texturing]

    style G fill:#5e6ce0,stroke:#3949ab,color:#fff

importMiddlebury is an offline-dataset ingest helper used outside the hot path.

The 12 binaries

aliceVision_cameraInit

Builds the initial SfMData (cameraInit.sfm) from a directory of images. Reads EXIF for focal length + sensor; looks up cameraSensors.db for the physical sensor size.

Input Output Notes
Image folder cameraInit.sfm (JSON SfMData) EXIF-driven focal-length init

aliceVision_featureExtraction

Per-view SIFT feature detection + description. CPU-only on Apple Silicon (vlsift) — the GPU SIFT path requires CUDA.

Input Output Notes
cameraInit.sfm features/ (per-view .feat + .desc) ~20K SIFT features/view typical

aliceVision_imageMatching

Selects which image pairs to attempt matching. Modes: Exhaustive, SequentialAndVocabularyTree, etc.

Input Output Notes
cameraInit.sfm, features imageMatches.txt exhaustive for small N (3 imgs → 3 pairs)

Target rename

The upstream imageMatching library and executable both have target name aliceVision_imageMatching — CMake forbids the collision. Our build renames the executable target to aliceVision_imageMatching_bin with OUTPUT_NAME aliceVision_imageMatching so the on-disk filename matches Meshroom's expectation (see memory/mental_note.md §8d).

aliceVision_featureMatching

Per-pair geometric matching + verification (RANSAC against fundamental / essential matrix model).

Input Output Notes
cameraInit.sfm, features, image pairs matches/ ~11.5K geometric matches on Monstree mini3

aliceVision_incrementalSfM

Bundle-adjusted Structure-from-Motion. Computes cameras + sparse 3D landmarks.

Input Output Notes
matches sfm.sfm 3 cameras + 7.4K landmarks on Monstree mini3

aliceVision_prepareDenseScene

Per-view undistortion. Produces .exr images aligned with the SfM calibration.

Input Output Notes
sfm.sfm dense/ (one EXR per view) EXR allows HDR intermediate

aliceVision_depthMapEstimation

The Metal-backed step. SGM + Refine + Optimize multi-view stereo. Calls through the 15 cuda_* adapter forwarders into our MSL kernels.

Input Output Notes
sfm.sfm + dense/ depthmaps/<viewID>_depthMap.exr + _simMap.exr ~12 s / view at 4032×3024 on M4 (S43); ~7.8 s after S44+S45 optimizations

The -2 value in _depthMap.exr is the alpha-mask sentinel for image-border pixels.

aliceVision_depthMapFiltering

Per-pixel depth refinement (consistency check, hole fill, normal-driven smoothing).

Input Output Notes
depthmaps/ depthmap_filtered/ ~1 s on the Monstree mini3 depth maps

aliceVision_meshing

Depth maps → 3D mesh via Delaunay-triangulated fuseCut + GraphCut visibility.

Input Output Notes
sfm.sfm, depthmap_filtered/ dense.sfm, mesh.obj 7.8K verts / 15.4K faces on Monstree mini3

Output format

--output foo.abc errors with "AliceVision is built without Alembic support." — macOS doesn't link Alembic. Use .sfm (dense SfMData JSON) or .ply instead. Source: memory/mental_note.md §8h-i.

aliceVision_meshFiltering

Laplacian smoothing + non-manifold cleanup.

Input Output Notes
mesh.obj smoothed mesh.obj 7.7K verts post-smoothing on Monstree mini3

aliceVision_texturing

UV unwrap + texture atlas baking. Per-camera reprojection with multi-band frequency contribution.

Input Output Notes
dense .sfm + mesh texturedMesh.obj + .mtl + PNG atlas 8192² PNG atlas on Monstree mini3 (~192 MB)

Two flags you'll forget

  1. Input is the dense .sfm (from meshing's --output), not the .ply point cloud.
  2. Default --colorMappingFileType=none skips actual baking — pass png / jpg / tif / exr to bake. Source: memory/mental_note.md §8h-ii.

aliceVision_importMiddlebury

Helper to ingest Middlebury MVS dataset .par files (calibrated camera matrices) into an SfMData. Note that Middlebury .par files have no landmarks; you still need the SfM cascade to produce them — see memory/mental_note.md §7g.

Input Output Notes
Middlebury .par + images cameraInit.sfm cameras + poses only (no landmarks)

Runtime environment

All 12 binaries expect:

Variable Purpose
ALICEVISION_ROOT Path to install prefix; share/aliceVision/{config.ocio,cameraSensors.db,luts/} resolved from here.
ALICEVISION_BIN_PATH (Meshroom-only) where Meshroom finds the binaries.
default.metallib (next to binary) Loaded via @executable_path; auto-staged by CMake.

Set via scripts/run_meshroom.sh for Meshroom-driven runs. See User → Running the pipeline for direct invocation examples.

Python-only nodes (not binaries)

SegmentationBiRefNet

Listed here for completeness

SegmentationBiRefNet is a Python Meshroom node, not an aliceVision_* CLI binary. It executes in-process via rembg + ONNX Runtime (CoreML EP) — no CMake target, no default.metallib dependency, no cuda_* adapter forwarder.

AI-powered foreground/background segmentation. Produces per-view masks consumed by downstream nodes (DepthMap, Meshing, Texturing).

Input Output Notes
CameraInit.output (image list) {nodeCacheFolder}/masks/{imageStem}_mask.png BiRefNet ONNX → CoreML (ANE + GPU + CPU). See Segmentation reference for the full parameter list.

Full docs:

Phase 14.1–14.9 additions (the other 49 binaries)

These were added across nine focused phases between 2026-05-23 and 2026-05-24 to bring template coverage from 4/25 to 25/25.

Modern SfM stack (Phase 14.1, 6 binaries)

aliceVision_sfmBootstrapping, aliceVision_sfmExpanding, aliceVision_relativePoseEstimating, aliceVision_sfmTriangulation, aliceVision_tracksBuilding, aliceVision_meshDecimate.

Used by photogrammetry.mg, cameraTracking*.mg. Needed a CMake-time relativePoses.hpp inline patch (ODR fix — upstream defines two tag_invoke free functions in the header without inline).

HDR pipeline (Phase 14.2, 3 binaries)

aliceVision_LdrToHdrSampling, aliceVision_LdrToHdrCalibration, aliceVision_LdrToHdrMerge. New aliceVision_hdr sublib via add_subdirectory. Drives hdrFusion.mg.

Panorama (Phase 14.3, 8 binaries + bonus)

aliceVision_panoramaInit, panoramaPrepareImages, panoramaWarping, panoramaEstimation, panoramaCompositing, panoramaMerging, panoramaSeams, panoramaPostProcessing. Plus bonus aliceVision_sfmTransform which unblocked panoramaHdr.mg + panoramaFisheyeHdr.mg. New aliceVision_panorama sublib.

Photometric stereo (Phase 14.4)

  • a (3 binaries) — aliceVision_lightingCalibration, aliceVision_photometricStereo (binary target name _bin + OUTPUT_NAME to avoid target-name collision with the static lib), bonus aliceVision_sfmTransfer. New aliceVision_photometricStereo + aliceVision_lightingEstimation sublibs. OpenCV find_package added.
  • b (1 binary, CoreML) — aliceVision_sphereDetection is a Mac-native CoreML port replacing upstream's ONNX-Runtime sphereDetection. Wraps ai-models/yolov8n.mlpackage via src/sphere_detection/. Runs the full graph on ANE (3× faster than GPU).

Utility + lidar (Phase 14.5, 21 binaries)

Camera tracking utilities: applyCalibration, checkerboardDetection, colorCheckerCorrection, colorCheckerDetection, convertDistortion, convertSfMFormat, depthmapTracksInjecting, distortionCalibration, exportDistortion, exportImages, geometricFilterEstimating, imageProcessing (_bin rename), intrinsicsTransforming, keyframeSelection, nodalSfM, sfmColorizing, sfmMerge, tracksMerging.

LIDAR: lidarDecimating, lidarMerging, lidarMeshing.

New sublibs: aliceVision_calibration, aliceVision_keyframe, aliceVision_imageProcessing (each PUBLIC_LINKS OpenCV).

Alembic export (Phase 14.6, 2 binaries)

aliceVision_exportAlembic, aliceVision_exportAnimatedCamera. Required flipping ALICEVISION_HAVE_ALEMBIC=1 BEFORE add_subdirectory(sfmDataIO) so the existing upstream sublib's own if(...) block folded AlembicExporter/Importer in.

Unlocks 8 cameraTracking-family templates.

Mac-port-native (Phase 14.7–14.9, 3 binaries)

For 2026.1.0 features upstream defers to external Python pipelines (no C++ source in our snapshot). Mac-port-native implementations:

Binary Phase Implementation
aliceVision_starListing 14.7 Algorithmic (star-topology image pair list). src/native_binaries/main_starListing.cpp.
aliceVision_matchMasking 14.9 CoreML TinyRoMa (replaces Phase 14.7 pass-through). Runs the dense matcher on each pair; writes per-pair _coarse_flow.exr, _fine_flow.exr, _coarse_certainty.exr, _fine_certainty.exr. Optional mask-aware filtering. src/roma/.
aliceVision_moGe 14.8 CoreML MoGe-2 (replaces Phase 14.7 stub). Per-view depth + normals at 504×672. src/moge/.

See ai-models/README.md for the per-model CoreML integration recipe + ANE outcome matrix.