Meshroom Manual Readthedocs Io en v19.01.45
Meshroom Manual Readthedocs Io en v19.01.45
Meshroom Manual Readthedocs Io en v19.01.45
Release 0.1
1 Manual 1
2 Install 3
4 Simple import 11
5 3D Viewer 13
7 Start Reconstruction 17
8 Augment Reconstruction 19
9 Live Reconstruction 23
10 External Reconstruction 25
12 Test Meshroom 29
14 Connect Nodes 33
16 Supported Formats 59
17 Tutorials 61
18 Capturing 75
19 More 77
i
21 References 95
22 Glossary 97
23 About 99
Index 103
ii
CHAPTER 1
Manual
1
Meshroom, Release 0.1
2 Chapter 1. Manual
CHAPTER 2
Install
2.1 Requirements
2.3 Windows
1. Download Meshroom
2. extract ZIP to a folder of yur choice
3
Meshroom, Release 0.1
2.4 Linux
4 Chapter 2. Install
Meshroom, Release 0.1
On Ubuntu, you may have conflicts between native drivers and mesa drivers. In that case, you need to force usage of
native drivers by adding them to the
LD_LIBRARY_PATH:
LD_LIBRARY_PATH=/usr/lib/nvidia-340 PYTHONPATH=$PWD python meshroom/ui
You may need to adjust the folder /usr/lib/nvidia-340 with the correct driver version.
Do not use foreign characters in path names (latin characters only).
Launch a 3D reconstruction in command line
Linux: PYTHONPATH=$PWD
python bin/meshroom_photogrammetry –input INPUT_IMAGES_FOLDER –output OUTPUT_FOLDER
German: http://paravel.org/blog/2018/12/10/how-to-photogrammetry-mit-meshroom-unter-ubuntu-16-04
2.5 OSX
2.5. OSX 5
Meshroom, Release 0.1
First off, your Mac will currently need an nVidia GPU with a CUDA compute capability of 2.0 or greater. This is
probably a pretty small portion of all Macs sold, but you can check your GPU by looking in “About This Mac” from
the Apple icon in the top left corner of the screen, under “Graphics”. If you have an nVidia GPU listed there, you can
check its compute capability on the nVidia CUDA GPUs page .
Second, you’re going to need to install the latest CUDA toolkit . As of this writing, that’s CUDA 9.2, which is only
officially compatible with OS X 10.13 (High Sierra), so you may also need to upgrade to the latest version of High
Sierra if you haven’t already. Alongside this I would also suggest installing the latest nVidia CUDA GPU webdriver,
which as of this writing is 387.10.10.10.40.105 .
Third, CUDA 9.2 is only ‘ compatible with the version of <https://docs.nvidia.com/cuda/
cuda-installation-guide-mac-os-x/index.html>‘_ clang ‘ distributed with Xcode 9.2 <https://docs.nvidia.com/
cuda/cuda-installation-guide-mac-os-x/index.html>‘_ , and will refuse to compile against anything else. You may
have an older or newer version of Xcode installed. As of this writing, if you fully update Xcode within a fully
updated OS X install, you’ll have Xcode 9.4.1. To get back to Xcode 9.2, what you can do is go to Apple’s Developer
Downloads page (for which you’ll need a free Apple developer account), then search for “Xcode 9.2”, then install the
Command Line Tools for Xcode 9.2 package for your OS version. After installing, run
sudo xcode-select –switch /Library/Developer/CommandLineTools and then verify that
clang –version * * shows Apple LLVM version 9.0.0
Once you’ve done all this, you can verify a working CUDA install by going to /Developer/NVIDIA/CUDA-
9.2/samples/1_Utilities/deviceQuery and running sudo make && ./deviceQuery , which should output your GPU in-
formation. If it doesn’t build correctly, or deviceQuery errors or doesn’t list your GPU, you may need to look over the
steps above and check that everything is up to date (you can also check the CUDA panel in System Preferences). ..
image:: homebrew-inst.jpg
2.5.2 Installation
If you’ve followed all the above setup instructions and requirements, installing the AliceVision libraries/framework
should be as easy as:
brew install ryanfb/alicevision/alicevision
I haven’t yet created a Homebrew formula for the Meshroom package itself , as it’s all Python and doesn’t seem par-
ticularly difficult to install/use once AliceVision is installed and working correctly. Just follow the install instructions
there (for my specific Python configuration/installation I used pip
3 instead of pip and python3 instead of python ):
git clone –recursive git://github.com/alicevision/meshroom
cd meshroom
pip install -r requirements.txt
You can report an issue on https://github.com/ryanfb/homebrew-alicevision/issues
One gotcha I ran into is that the CUDA-linked AliceVision binaries invoked by Meshroom don’t automatically find the
CUDA libraries on the DYLD_LIBRARY_PATH , and setting the DYLD_LIBRARY_PATH from the shell launching
6 Chapter 2. Install
Meshroom, Release 0.1
Meshroom doesn’t seem to get the variable passed into the shell environment Meshroom uses to spawn commands.
Without this, you’ll get an error like:
dyld: Library not loaded: @rpath/libcudart.9.2.dylib
Referenced from: /usr/local/bin/aliceVision_depthMapEstimation
Reason: image not found
In order to get around this, you can symlink the CUDA libraries into /usr/local/lib (most of the other workarounds I
found for permanently modifying the DYLD_LIBRARY_PATH seemed more confusing or fragile than this simpler
approach): 1
for i in /Developer/NVIDIA/CUDA-9.2/lib/.a /Developer/NVIDIA/CUDA-9.2/lib/.dylib; do ln -sv “$i”
“/usr/local/lib/$(basename “$i”)”; done
You can undo/uninstall this with:
for i in /Developer/NVIDIA/CUDA-9.2/lib/.a /Developer/NVIDIA/CUDA-9.2/lib/.dylib; do rm -v
“/usr/local/lib/$(basename “$i”)”; done
You may also want to download the voctree dataset:
curl ‘https://gitlab.com/alicevision/trainedVocabularyTreeData/raw/master/vlfeat_K80L3.SIFT.tree’ -o
/usr/local/Cellar/alicevision/2.0.0/share/aliceVision/vlfeat_K80L3.SIFT.tree
Then launch with:
ALICEVISION_SENSOR_DB=/usr/local/Cellar/alicevision/2.0.0/share/aliceVision/cameraSensors.db ALICE-
VISION_VOCTREE=/usr/local/Cellar/alicevision/2.0.0/share/aliceVision/vlfeat_K80L3.SIFT.tree PYTHON-
PATH=$PWD python meshroom/ui
Import some photos, click “Start”, wait a while, and hopefully you should end up with a reconstructed and textured
mesh . By default, the output will be in MeshroomCache/Texturing/ (relative to where you saved the project file).
When you launch Meshroom without sudo, the temp path will be something like this:
2.5. OSX 7
Meshroom, Release 0.1
2.6 Docker
(WIP)
official: docker pull alicevision/meshroom https://hub.docker.com/r/alicevision/meshroom
Docker file on Github: https://github.com/alicevision/meshroom/blob/master/Dockerfile
Other:
https://hub.docker.com/r/derfetzer/meshroom/ https://hub.docker.com/r/fschwaiger/meshroom
Link to CUDA and Docker https://stackoverflow.com/questions/25185405/using-gpu-from-a-docker-container
8 Chapter 2. Install
CHAPTER 3
Images Pane
3D Viewer Pane
Cache Folder File Path (where temp files and final results are stored)
9
Meshroom, Release 0.1
You can grab a Pane border and move it to change the pane size.
Simple import
Drag-n-drop your images or your image folder into the Images pane on the left hand side.
You can preview the images in the Image Viewer pane. To display the image metadata click the (i) icon in the bottom
right corner. For images with embedded GPS information an additional openstreetmap frame will be displayed.
11
Meshroom, Release 0.1
Note: If your images won’t appear in the Images pane after you imported them, your camera was not recognized
correctly. Make sure the EXIF data contains all relevant camera information. If the import still fails, your camera is
not in the database or your image files are not valid.
3D Viewer
The 3D Viewer will preview the SfM Pointcloud, cameras and the Mesh preview. You can use your mouse or the
rotate/scale toolbar on the left. You can hold Shift to pan. Press F to reset the view. Double-click to create a new
rotation center for the Mesh. To display the final model, a button will appear on the bottom side to load the mesh
(Load model). Uncheck the SfM layer for a better view. To refit the 3D-model to the new dimensions of the pane if
you changed its size, right-click to display a menu with refitting options.
13
Meshroom, Release 0.1
By default StructureFromMotion and Texturing results will be added to your Scene layers. You can add the out-
puts of other node variations to your Scene in the 3D Viewer by double clicking on the nodes. Supported nodes:
StructureFromMotion, Texturing, MeshDecimate, MeshDenoise, MeshResampling
3D Model The final 3D-Model will be saved in Project Folder →MeshroomCache → Texturing
By default it will be saved in the OBJ format. You can change it in the node settings.
Note: At the moment Meshroom does not support model realignment, so the model can be orientated upside down
relative to the grid. You can change the orientation in another software like Meshlab.
14 Chapter 5. 3D Viewer
CHAPTER 6
This PR introduces the notion of “advanced” parameters on nodes. The goal is to separate experimen-
tal/debug/advanced from end-user attributes. On the UI side, the AttributeEditor has been redesigned and now provides
an additional option to show/hide those advanced parameters.
15
Meshroom, Release 0.1
Start Reconstruction
Click the green Start button to start processing. To stop/pause click the Stop button. The progress will be kept.
There are two progress bars: the line below the menu bar indicating the overall progress and the other in the Graph
Editor within the nodes. To get a detailed progress log, open the CommandLine window or click on the node you
are interested in and go to the Log tab in the properties pane of the Graph Editor.
You can open the (Your-Project-Folder) -> MeshroomCache to see the output of each node. (Shortcut: Icon and path
at the bottom left side of the main window)
A node folder contains the output of the node. By default Meshroom uses a unique id to name the output folders to
prevent overwriting data and already computed results of the project can be reused.
Example: You are not satisfied with your first result and make changes to the StructureFromMotion node. The new
output will be placed under a different name inside the StructureFromMotion Folder.
You can change the name of the output folders of your nodes by clicking on the node and changing the Output
Folder name in the Attributes tab of the Graph Editor Properties pane.
17
Meshroom, Release 0.1
Augment Reconstruction
You can drag-n-drop additional images into the lower part of the Images Pane, called Augment Reconstruction. For
each batch of images, a new Group will be created in the Images Pane. You can drop successive batches of N images
in the Images Pane. for each batch of images the graph will branch.
You can use this method for complex scenes with multiple objects
Note: The groups will be merged using the ImageMatchingMultiSfM node. Read the node description for details
19
Meshroom, Release 0.1
21
Meshroom, Release 0.1
Live Reconstruction
Live reconstruction is meant to be used along with a camera that can transfer images to a computer while shooting
(using wifi, a wifi sd-card or Tethering). Meshroom can watch a folder for new images and successively augment
previous SfM (point clouds + cameras) after each {Min. Images} per Step. This allows to get an iterative preview
during shooting, e.g to see which areas of the dataset requires more coverage.
To enable Live Reconstruction go to the menu bar View -> Live Reconstruction A new Live Reconstruction pane
will appear under the Images pane.
For each new import, a new Image Group inside the Images pane will be created. Also the Graph Editor updates
the graph, adding nodes to process the newly added images and add them to the pipeline.
Select the Image Folder to watch and the minimum of new images folder to be imported per step. Click Start in the
Live Reconstruction pane to start monitoring the selected folder for new files. You should then see in the graph one
branch (from CameraInit to StructureFromMotion) for each batch of images. 1 The reconstruction process will
stop at the last processed StructureFromMotion node and will not automatically go through the rest of the default
pipeline. This is for practical reasons. The point cloud will update in real time with newly added images. Computing
the mesh for every new image batch is not effective.
Once you complete the image capturing process, click Stop and disconnect the PrepareDenseScene node from the
first StructureFromMotion node and connect it with the last StructureFromMotion node.
23
Meshroom, Release 0.1
Note: The groups will be merged using the ImageMatchingMultiSfM node. Read the node description for details.
External Reconstruction
Use this option when you compute externally after submission to a render farm from meshroom. (need to have access
to a renderfarm and need the corresponding submitter).
This way, you can make use of external computing power. If you can not compute GPU nodes locally (no cuda) you
can still submit them.
Available submitters:
25
Meshroom, Release 0.1
27
Meshroom, Release 0.1
Test Meshroom
For your first reconstruction in Meshroom, download the Monstree Image Dataset https://github.com/
alicevision/dataset_monstree. You can preview the Monstree model on Sketchfab https://sketchfab.com/models/
92468cb8a14a42f39c6ab93d24c55926.
The Monstree dataset is known to work, so there should be no errors or problems during the reconstruction. This
might be different when using your own image dataset.
Import the images in Meshroom by dropping them in the Images pane. There are different folders in the Monstree
dataset: full (all images), mini6 (6 images) and mini3 (3 images) to test out.
You can preview selected images in the Image Viewer pane. To display the image metadata click the (i) icon in the
bottom right corner. For images with embedded GPS information an additional openstreetmap frame will be displayed.
In the Graph Editor you can see the ready-to-use default pipeline.
29
Meshroom, Release 0.1
The Graph Editor contains the processing nodes of your pipeline. For this project you do not need to change anything!
In fact, for many projects the default pipeline delivers good results
You can zoom in or restructure the nodes. You can hold Shift to pan using the mouse. To insert new nodes right-click
in the Graph Editor pane. For the Graph Editor use the buttons on the bottom left side of the pane to (re)order.
Before you start the reconstruction, save the project to the Monstree folder. (File ? Save as) (The HDD should have
enough free space.)
You can calculate with 30sec. per image on a computer with i7@2,9GHz, GTX1070 8GB, 32GB Ram.
Performance: % of overall processing time with default pipeline: ~38% DepthMap / ~24% Meshing
With the 2019.1.0 release, the reconstruction time has been reduced by ~30% compared to the 2018.1 release. The
Cache Folder file size has been reduced by 20% Tested with the Monstree Dataset: (comparing only computing time,
not quality) Computing time in seconds: (total MR2018 260s / MR2019 185s)
https://web.archive.org/web/20181010161448/https://scanbox.xyz/blog/alicevision-opensource-photogrammetry/
31
Meshroom, Release 0.1
For a Full Pipeline Evaluation including the “Tanks and Temples” evaluation benchmark read D5.4: Deliver 3D
reconstruction benchmarks with dataset available on https://cordis.europa.eu/project/rcn/205980/results/en in Doc-
uments, reports.
Connect Nodes
The node connections of the default Graph can be difficult to understand. The following images illustrate how the
nodes are connected.
This image illustrates the default graph with node connections on the origin nodes:
33
Meshroom, Release 0.1
node_reference/connect-nodes/draft-meshingnode-graph.jpg
node_reference/connect-nodes/draft-meshingnode-graph-color.jpg
Description
Note: This node requires AliceVision compiled with opencv. Not included in the MR 2019.1 binary.
The internal camera parameters can be calibrated from multiple views of a checkerboard. This allows to retrieve focal
length, principal point and distortion parameters. A detailed explanation is presented in [opencvCameraCalibration].
[opencvCameraCalibration] http://docs.opencv.org/3.0-beta/doc/tutorials/calib3d/camera_calibration/camera_
calibration.html
35
Meshroom, Release 0.1
Table 1: settings
Name Description
Input Input images in one of the following form: – folder containing images – image sequence like
“/path/to/seq.@.jpg” – video file
Pattern Type of pattern (camera calibration patterns) - CHESSBOARD - CIRCLES - ASYMMET-
RIC_CIRCLES - ASYMMETRIC_CCTAG
Size (Size of the Pattern) - Number of inner corners per one of board dimension like Width Height
Square Size Size of the grid’s square cells (0-100mm)
Nb Distortion Coef Number of distortion coefficient (0-5)
Max Frames Maximal number of frames to extract from the video file (0-5)
Calib Grid Size Define the number of cells per edge (0-50)
Max Calib Frames Maximal number of frames to use to calibrate from the selected frames (0-1000)
Min Input Frames Minimal number of frames to limit the refinement loop (0-100)
Max Total Average Max Total Average Error (0-1)
Error
Debug Rejected Folder to export delete images during the refinement loop
Img Folder
Debug Selected Folder to export debug images
Img Folder
Output Output filename for intrinsic [and extrinsic] parameters (default filename cameraCalibra-
tion.cal)
15.2 CameraInit
Description
-load image metadata and sensor information You can mix multiple cameras and focal lengths. The CameraInit will
create groups of intrinsics based on the images metadata. It is still good to have multiple images with the same camera
and same focal lengths as it adds constraints on the internal cameras parameters. But you can combine multiple groups
of images, it will not decrease the quality of the final model.1
Note: In some cases, some image(s) have no serial number to identify the camera/lens device. This makes it impos-
sible to correctly group the images by device if you have used multiple identical (same model) camera devices. The
reconstruction will assume that only one device has been used, so if 2 images share the same focal length approxima-
tion they will share the same internal camera parameters. If you want to use multiple cameras, add a corresponding
serialnumber to the EXIF data.
Table 2: settings
Name Description
Viewpoints Input viewpoints
(1 Element for each loaded image)
ID Pose ID Image Path Intrinsic:
Internal Camera Parameters (Intrinsic
ID) Rig (-1 - 200) Rig Sub-Pose:
Rig Sub-Pose Parameters (-1 - 200)
Image Metadata: (list of metadata
elements)
Intrinsic Camera Intrinsics (1 Element for each loaded image) ID Initial Focal
Length: Initial Guess on the Focal Length Focal Length:
Known/Calibrated Focal Length Camera Type: pin-
hole’, ‘radial1’, ‘radial3’, ‘brown’, ‘fisheye4’ #Make:
Camera Make (not included in this build, commented
out) #Model: Camera Model #Sensor Width: Camera
Sensor Width Width: Image Width (0-10000) Height:
Image Height (0-10000) Serial Number: Device Serial
Number (camera and lens combined) Principal Point:
X (0-10000) Y(0-10000) DistortionParams: Distortion
Parameters Locked(True/False): If the camera has been
calibrated, the internal camera parameters (intrinsics)
can be locked. It should improve robustness and
speedup the reconstruction.
Sensor Database Camera sensor width database path
Default Field Of View Empirical value for the field of view in degree 45° (0°-
180°)
Verbose Level verbosity level (fatal, error, warning, info, debug, trace)
Output SfMData File . . . /cameraInit.sfm
Notes
Issue: structure from motion reconstruction appears distorted, and has failed to aligned some groups of cameras when
loading images without focallength
Solution: Keep the ” Focal Length” init value but set the “Initial Focal Length” to -1 if you are not sure of the value.
https://github.com/alicevision/meshroom/issues/434
Description
Based on the SfM results, we can perform camera localization and retrieve the motion of an animated camera in the
scene of the 3D reconstruction. This is very useful for doing texture reprojection in other software as part of a texture
clean up pipeline. Could also be used to leverage Meshroom as a 3D camera tracker as part of a VFX pipeline
https://alicevision.github.io/#photogrammetry/localization
Table 3: settings
Name Description
SfM Data The sfm_data.json kind of file generated by AliceVision
Media File The folder path or the filename for the media to track
Visual Debug If a folder is provided it enables visual debug and saves all the debugging info in that folder
Folder
Descriptor Path Folder containing the descriptors for all the images (ie the .desc.)
Match Desc Describer types to use for the matching: sift’, ‘sift_float’, ‘sift_upright’, ‘akaze’, ‘akaze_liop’,
Types ‘akaze_mldb’, ‘cctag3’, ‘cctag4’, ‘sift_ocv’, ‘akaze_ocv
Preset Preset for the feature extractor when localizing a new image (low, medium, normal, high, ultra)
Resection Esti- The type of /sac framework to use for resection (acransac, loransac)
mator
Matching Esti- The type of /sac framework to use for matching (acransac, loransac)
mator
Calibration Calibration file
Refine Intrin- Enable/Disable camera intrinsics refinement for each localized image
sics
Reprojection Maximum reprojection error (in pixels) allowed for resectioning. If set to 0 it lets the ACRansac
Error select an optimal value (0.1 - 50)
Nb Image [voctree] Number of images to retrieve in database (1 - 1000)
Match
Max Results [voctree] For algorithm AllResults, it stops the image matching when this number of matched
images is reached. If 0 it is ignored (1 - 100)
Commonviews [voctree] Number of minimum images in which a point must be seen to be used in cluster
tracking (2 - 50)
Voctree [voctree] Filename for the vocabulary tree
Voctree [voctree] Filename for the vocabulary tree weights
Weights
Algorithm [voctree] Algorithm type: (FirstBest, AllResults)
Matching Error [voctree] Maximum matching error (in pixels) allowed for image matching with geometric ver-
ification. If set to 0 it lets the ACRansac select an optimal value (0 - 50)
Nb Frame [voctree] Number of previous frame of the sequence to use for matching (0 = Disable) (0 - 100)
Buffer Match-
ing
Robust Match- [voctree] Enable/Disable the robust matching between query and database images, all putative
ing matches will be considered
N Nearest Key [cctag] Number of images to retrieve in the database Parameters specific for final (optional)
Frames bundle adjustment optimization of the sequence: (1-100)
Global Bundle [bundle adjustment] If –refineIntrinsics is not set, this option allows to run a final global bundle
adjustment to refine the scene
No Distortion [bundle adjustment] It does not take into account distortion during the BA, it consider the dis-
tortion coefficients all equal to 0
No BA Refine [bundle adjustment] It does not refine intrinsics during BA
Intrinsics
Min Point Visi- [bundle adjustment] Minimum number of observation that a point must have in order to be
bility considered for bundle adjustment (2-50)
Output Alem- Filename for the SfMData export file (where camera poses will be stored)
bic desc.Node.internalFolder + ‘trackedCameras.abc
Output JSON Filename for the localization results as .json desc.Node.internalFolder + ‘trackedCameras.json
Description
If a rig of cameras is used, we can perform the rig calibration. We localize cameras individually on the whole sequence.
Then we use all valid poses to compute the relative poses between cameras of the rig and choose the more stable value
across the images. Then we initialize the rig relative pose with this value and perform a global Bundle Adjustment
on all the cameras of the rig. When the rig is calibrated, we can use it to directly localize the rig pose from the
synchronized multi-cameras system with [Kneip2014] approaches.
..The rig calibration find the relative poses between all cameras used. It takes a point cloud as input and can use both
CCTag and SIFT features for localization. The implication is that all cameras must see features (either SIFT or CCTag)
that are part of the point cloud, but they do not have to observe overlapping regions. (See:POPART: Previz for Onset
Production Adaptive Realtime Tracking)
“Given the position of the tracked reference frame relative to the motion capture system and the optical reference
frames it is possible to retrieve the transformation between the tracked and the optical reference frames”1 “In practice,
it is particularly difficult to make the tracked frame coincident with the camera optical frame, thus a calibration
procedure is needed to estimate this transformation and achieve the millimetric accuracy” [Chiodini et al. 2018]
[Chiodini et al. 2018] Chiodini, Sebastiano & Pertile, Marco & Giubilato, Riccardo & Salvioli, Federico & Bar-
rera, Marco & Franceschetti, Paola & Debei, Stefano. (2018). Camera Rig Extrinsic Calibration Using a Motion
Capture System. 10.1109/MetroAeroSpace.2018.8453603. https://www.researchgate.net/publication/327513182_
Camera_Rig_Extrinsic_Calibration_Using_a_Motion_Capture_System
https://alicevision.github.io/#photogrammetry/localization
[Kneip2011] A Novel Parametrization of the Perspective-Three-Point Problem for a Direct Computation of Absolute
Camera Position and Orientation. L. Kneip, D. Scaramuzza, R. Siegwart. June 2011
[Kneip2013] Using Multi-Camera Systems in Robotics: Efficient Solutions to the NPnP ProblemL. Kneip, P. Furgale,
R. Siegwart. May 2013
[Kneip2014] OpenGV: A unified and generalized approach to real-time calibrated geometric vision, L. Kneip, P.
Furgale. May 2014.
[Kneip2014] Efficient Computation of Relative Pose for Multi-Camera Systems. L. Kneip, H. Li. June 2014
Table 4: settings
NameDescription
SfM The sfmData file
Data
Me- The path to the video file the folder of the im-
dia age sequence or a text
Path file (one image path per
line) for each camera of
the rig (eg. –media-
path /path/to/cam1.mov
/path/to/cam2.mov)
Cam- The intrinsics calibration
era file for each camera of
In- the rig. (eg. –cameraIn-
trin- trinsics /path/to/calib1.txt
sics /path/to/calib2.txt)
Ex- Filename for the alembic
port file containing the rig
poses with the 3D points.
It also saves a file for
each camera named
‘filename.cam##.abc
(trackedcameras.abc)
De- Folder containing the
scrip- .desc
tor
Path
Match The describer types to use ‘sift_float’ ‘sift_upright’
‘akaze’‘akaze_liop’
‘akaze_mldb’
‘cc- ‘cc- ‘sift_ocv’
‘akaze_ocv’‘‘
De- for the matching ‘‘’sift’ tag3’ tag4’
scriber
Types
Pre- Preset for the feature ex- medium nor- high ul-
set tractor when localizing a mal tra)
new image (low
Re- The type of /sac frame- loransac)
sec- work to use for resection
tion (acransac
Es-
ti-
ma-
tor
Match-The type of /sac frame- loransac)
ing work to use for matching
Es- (acransac
ti-
ma-
tor
Re- Enable/Disable camera
fine intrinsics refinement for
In- each localized image
trin-
sics
Re- Maximum reprojection
pro- error (in pixels) allowed
jec- for resectioning. If set
40tion to 0 it lets the ACRansac Chapter 15. Complete Node List
Er- select an optimal value.
ror (0 - 10)
Max Maximum number of
Meshroom, Release 0.1
Voctree Weights: http://www.ipol.im/pub/art/2018/199/ voctree (optional): For larger datasets (>200 images), greatly
improves image matching performances. It can be downloaded here. https://github.com/fragofer/voctree You need to
specify the path to vlfeat_K80L3.SIFT.tree in Voctree.
Description
This node retrieves the transformation between the tracked and the optical reference frames.(?) https://alicevision.
github.io/#photogrammetry/localization
Table 5: settings
NameDescription
SfM The sfmData file
Data
Me- The path to the video file the folder of the im-
dia age sequence or a text
Path file (one image path per
line) for each camera of
the rig (eg. –media-
path /path/to/cam1.mov
/path/to/cam2.mov)
Rig The file containing the
Cal- calibration data for the rig
i- (subposes)
bra-
tion
File
Cam- The intrinsics calibration
era file for each camera of
In- the rig. (eg. –cameraIn-
trin- trinsics /path/to/calib1.txt
sics /path/to/calib2.txt)
De- Folder containing the
scrip- .desc
tor
Path
Match The describer types to use ‘sift_float’ ‘sift_upright’
‘akaze’‘akaze_liop’
‘akaze_mldb’
‘cc- ‘cc- ‘sift_ocv’
‘akaze_ocv’)‘‘
De- for the matching ‘‘(sift’ tag3’ tag4’
scriber
Types
Pre- Preset for the feature ex- medium nor- high ul-
set tractor when localizing a mal tra)‘‘
new image ‘‘(low
Re- The type of /sac frame- loransac)‘‘
sec- work to use for resection
tion ‘‘(acransac
Es-
ti-
ma-
tor
Match-The type of /sac frame- loransac)‘‘
ing work to use for matching
Es- ‘‘(acransac
ti-
ma-
tor
Re- Enable/Disable camera
fine intrinsics refinement for
In- each localized image
trin-
sics
Re- Maximum reprojection
pro- error (in pixels) allowed
jec- for resectioning. If set
tion to 0 it lets the ACRansac
42Er- select an optimal value (0 Chapter 15. Complete Node List
ror - 10)
Use Enable/Disable the naive
Lo- method for rig localiza-
Meshroom, Release 0.1
15.6 ConvertSfMFormat
Description
• creates abc’, ‘sfm’, ‘json’, ‘ply’, ‘baf SfM File from SfMData file
Table 6: settings
Name Description
Input SfMData file
SfM SfM File Format ‘‘(output file ex- ‘sfm’ ‘json’ ‘ply’ ‘baf)‘‘
File tension: abc’
Format
De- Describer types to keep.‘‘’sift’ ‘sift_float’
‘sift_upright’
‘akaze’‘akaze_liop’
‘akaze_mldb’
‘cc- ‘cc- ‘sift_ocv’
‘akaze_ocv’‘‘
scriber tag3’ tag4’
Types
Image Image id
id
Image image white list (uids or image
White paths).
List
Views Export views
Intrin- Export intrinsics
sics
Extrin- Export extrinsics
sics
Struc- Export structure
ture
Obser- Export observations
vations
Ver- verbosity level ‘‘(fatal er- warn- info de- trace)‘‘
bose ror ing bug
Level
Output Path to the output SfM Data
file. (desc.Node.internalFolder +
‘sfm.{fileExtension})
15.6. ConvertSfMFormat 43
Meshroom, Release 0.1
15.7 DepthMap
Description
Table 7: settings
Name Description
MVS SfMData file.
Configuration
File:
Images Folder Use images from a specific folder instead of those specify
in the SfMData file.Filename should be the image uid.
Downscale Image downscale factor ‘‘(1 2 4 8 16)‘‘
Min View Angle Minimum angle between two views. ‘‘(0.0 10.0 0.1)‘‘
Max View Angle Maximum angle between two views. ‘‘(10.0 120.0 1)‘‘
SGM: Nb Neigh- Semi Global Matching: Number of neighbour cameras (1
bour Cameras - 100)
SGM: WSH: Semi Half-size of the patch used to compute the similarity (1 -
Global Matching 20)
SGM: GammaC Semi Global Matching: GammaC Threshold (0 - 30)
SGM: GammaP Semi Global Matching: GammaP Threshold (0 - 30)
Refine: Number of (1 - 500)
samples
Refine: Number of (1 - 100)
Depths
Refine: Number of (1 - 500)
Iterations
Refine: Nb Neigh- Refine: Number of neighbour cameras. (1 - 20)
bour Cameras
Refine: WSH Refine: Half-size of the patch used to compute the simi-
larity. (1 - 20)
Refine: Sigma Refine: Sigma Threshold (0 - 30)
Refine: GammaC Refine: GammaC Threshold. (0 - 30)
Refine: GammaP Refine: GammaP threshold. (0 - 30)
Refine: Tc or Rc Use minimum pixel size of neighbour cameras (Tc) or cur-
pixel size rent camera pixel size (Rc)
Verbose Level verbosity level (fatal er- warn- info de- trace)
ror ing bug
Output Output folder for generated depth maps
default:
15.8 DepthMapFilter
Description
The original depth maps will not be entirely consistent. Certain depth maps will claim to see areas that are occluded
by other depth maps. The DepthMapFilter step isolates these areas and forces depth consistency.
Table 8: settings
Name Description
Input SfMData file
Depth Map Folder Input depth map folder
Number of Nearest Number of nearest cameras used for filtering 10 (0
Cameras - 20)
Min Consistent Cameras Min Number of Consistent Cameras 3 (0 - 10)
Min Consistent Cameras Min Number of Consistent Cameras for pixels with
Bad Similarity weak similarity value 4 (0 - 10)
Filtering Size in Pixels Filtering size in Pixels (0 - 10)
Filtering Size in Pixels Filtering size in pixels (0 - 10)
Bad Similarity
Verbose Level verbosity level (fatal er- warn- info de- trace)
ror ing bug
Output Output folder for generated depth maps
Min Consistent Cameras lower this value if the Meshing node has 0 depth samples input
View Output open output folder and view EXR files
15.9 ExportAnimatedCamera
Description
creates an Alembic animatedCamera.abc file from SFMData (e.g. for use in 3D Compositing software)
Table 9: settings
Name Description
Input SfMData file containing a complete
SfMData SfM
SfMData Filter A SfMData file use as filter
Export Undis- Export Undistorted Images value=True
torted Images
Undistort Image Image file format to use for undistorted images
*.jpg *.tif *.exr
Format ‘‘(*.jpg
(half))‘‘
Verbose Level Verbosity level ‘‘(fatal er- warn- info de- trace)‘‘
ror ing bug
Output filepath Output filepath for the alembic animated camera
Output Camera Output filename for the alembic animated camera
Filepath internalFolder + ‘camera.abc’
15.8. DepthMapFilter 45
Meshroom, Release 0.1
15.10 ExportMaya
Description
Mode for use with MeshroomMaya plugin.
The node “ExportMaya” exports the undistorted images. This node has nothing dedicated to Maya but was used to
import the data into our MeshroomMaya plugin. You can use the same to export to Blender.
15.11 FeatureExtraction
Description
15.12 FeatureMatching
Description
15.13 ImageMatching
Description
15.14 ImageMatchingMultiSfM
Description
This node can combine image matching between two input SfMData.
Used for Live Reconstructin and Augmentation
15.15 KeyframeSelection
Description Note: This is an experimental node for keyframe selection in a video, which removes too similar or too
blurry images. This node is not yet provided in the binaries as it introduces many dependencies. So if you built
it by yourself, you can test the KeyframeSelection node. It is not yet fully integrated into Meshroom, so you have
to manually drag&drop the exported frames to launch the reconstruction (instead of just adding a connection in the
graph) https://github.com/alicevision/meshroom/issues/232
15.16 MeshDecimate
Description
15.15. KeyframeSelection 49
Meshroom, Release 0.1
Simplify your mesh to reduce mesh size without changing visual appearance of the model.
or Meshing->MeshDecimate->MeshFiltering?
Comparison MeshDecimate and MeshResampling
Flip Normals
15.17 MeshDenoising
Description
Denoise your mesh Mesh models generated by 3D scanner always contain noise. It is necessary to remove the
noise from the meshes. Mesh denoising: remove noises, feature-preserving https://www.cs.cf.ac.uk/
meshfiltering/index_files/Doc/Random%20Walks%20for%20Mesh%20Denoising.ppt
15.17. MeshDenoising 51
Meshroom, Release 0.1
15.18 MeshFiltering
Description
Filter out unwanted elements of your mesh
Note: “Keep Only The Largest Mesh”. This is disabled by default in the 2019.1.0 release to avoid that the environment
is being meshed, but not the object of interest. The largest Mesh is in some cases the reconstructed background. When
15.18. MeshFiltering 53
Meshroom, Release 0.1
the object of interest is not connected to the large background mesh it will be removed. You should place your object
of interest on a well structured non transparent or reflecting surface (e.g. a newspaper).
15.19 MeshResampling
Description
Reducing number of faces while trying to keep overall shape, volume and boundaries You can specify a fixed, min,
max Vertices number.
This is different from MeshDecimate!
Resampling https://users.cg.tuwien.ac.at/stef/seminar/MeshResamplingMerge1901.pdf
Flip Normals
15.20 Meshing
Description
none
15.20. Meshing 55
Meshroom, Release 0.1
15.21 PrepareDenseScene
Description
• This node undistorts the images and generates EXR images
15.22 Publish
Description
• A copy of the Input files are placed in the Output Folder
Can be used to save SfM, Mesh or textured Model to a specific folder
15.23 SfMAlingnment
15.24 SfMTransform
Description
Apply a given transformation camera as the origin of the coordinate system with the SfMTransform node. You can
rescale the scene based on the bounding box of CCTAG markers.
15.23. SfMAlingnment 57
Meshroom, Release 0.1
15.25 StructureFromMotion
Description
none
15.26 Texturing
Description
Texturing creates UVs and projects the textures change quality and size/ file type of texture
Supported Formats
59
Meshroom, Release 0.1
NameRefer-Description
nce
Alem- Alem- cloud_and_poses Alembic is a format for storing information about animated
bic bic scenes after programmatic elements have been applied.
(.abc)
OBJ OBJ is a very strict ASCII format for encoding vertices points faces and
textures
first intro-
duced by
Wavefront
Technolo-
gies.
PLY PLY The Polygon File Format (or Stanford Triangle Format) has an ASCII represen-
tation and a binary representation. It is inspired by the OBJ format that allows
the definition of arbitrary properties for every point. This allows an implemen-
tation to add arbitrary information to points including accuracy information, but
not in any backward-compatible way. Camera information could be included
in comments.
SfM
.bin denseReconstruction: The bin format is only useful to get the visibility information of each vertex (no color
information)
.cal calibration file
.desc describer file
.EXR OpenEXR image format: for depth map images
.txt text file list to describer image parameters .ini A configuration file
.json describes the used image dataset
.baf (sfm) Bundle Adjustment File Export SfM data (Intrinsics/Poses/Landmarks)
Tutorials
17.1 Turntable
https://sketchfab.com/blogs/community/tutorial-meshroom-for-beginners
61
Meshroom, Release 0.1
17.3 Goal
In this tutorial, we will explain how to use Meshroom to automatically create 3D models from a set of photographs.
After specifying system requirements and installation, we will begin with some advice on image acquisition for pho-
togrammetry. We will then give an overview of Meshroom UI and cover the basics by creating a project and starting
the 3D reconstruction process. After that, we will see how the resulting mesh can be post-processed directly within
Meshroom by applying an automatic decimation operation, and go on to learn how to retexture a modified mesh. We
will sum up by showing how to use all this to work iteratively in Meshroom.
Finally, we will give some tips about uploading your 3D models to Sketchfab and conclude with useful links for further
information.
Meshroom software releases are self-contained portable packages. They are uploaded on the project’s GitHub page.
To use Meshroom on your computer, simply download the proper release for your OS (Windows and Linux are
supported), extract the archive and launch Meshroom executable.
Regarding hardware, an Nvidia GPU is required (with Compute Capability of at least 2.0) for the dense high quality
mesh generation. 32GB of RAM is recommended for the meshing, but you can adjust parameters if you don’t meet
this requirement.
Meshroom is released in open source under the permissive MPLv2 license, see Meshroom COPYING for more infor-
mation.
The shooting quality is the most important and challenging part of the process. It has dramatic impacts on the quality
of the final mesh.
The shooting is always a compromise to accomodate to the project’s goals and constraints: scene size, material prop-
erties, quality of the textures, shooting time, amount of light, varying light or objects, camera device’s quality and
settings.
The main goal is to have sharp images without motion blur and without depth blur. So you should use tripods or fast
shutter speed to avoid motion blur, reduce the aperture (high f-number) to have a large depth of field, and reduce the
ISO to minimize the noise.
For this first step, we will only use the high-level UI. Let’s save this new project on our disk using “File > Save As. . . ”.
All data computed by Meshroom will end up in a “MeshroomCache” folder next to this project file. Note that projects
are portable: you can move the “.mg” file and its “MeshroomCache” folder afterwards. The cache location is indicated
in the status bar, at the bottom of the window.
Next, we import images into this project by simply dropping them in the “Images” area – on the left-hand side.
Meshroom analyzes their metadata and sets up the scene.
Meshroom relies on a Camera Sensors Database to determine camera internal parameters and group them together. If
your images are missing metadata and/or were taken with a device unknown to Meshroom, an explicit warning will be
displayed explaining the issue. In all cases, the process will go on but results might be degraded.
Once this is done, we can press the “Start” button and wait for the computation to finish. The colored progress bar
helps follow the progress of each step in the process:
• green: has been computed
• orange: is being computed
• blue: is submitted for computation
• red: is in error
The generic photogrammetry pipeline can be seen as having two main steps:
• SfM: Structure-from-Motion (sparse reconstruction)
– Infers the rigid scene structure (3D points) with the pose (position and orientation) and internal calibration
of all cameras.
– The result is a set of calibrated cameras with a sparse point cloud (in Alembic file format).
• MVS: MultiView-Stereo (dense reconstruction)
– Uses the calibrated cameras from the Structure-from-Motion to generate a dense geometric surface.
– The final result is a textured mesh (in OBJ file format with the corresponding MTL and texture files).
As soon as the result of the “Structure-from-Motion” is available, it is automatically loaded by Meshroom. At this
point, we can see which cameras have been successfully reconstructed in the “Images” panel (with a green camera icon)
and visualize the 3D structure of the scene. We can also pick an image in the “Images” panel to see the corresponding
camera in the 3D Viewer and vice-versa.
There is no export step at the end of the process: the resulting files are already available on disk. You can right-click
on a media and select “Open Containing Folder” to retrieve them. By doing so on “Texturing”, we get access to the
folder containing the OBJ and texture files.
Let’s now see how the nodal system can be used to add a new process to this default pipeline. The goal of this step
will be to create a low-poly version of our model using automatic mesh decimation.
Let’s move to the “Graph Editor” and right click in the empty space to open the node creation menu. From there, we
select “MeshDecimate”: this creates a new node in the graph. Now, we need to give it the high-poly mesh as input.
Let’s create a connection by clicking and dragging from MeshFiltering.output to MeshDecimate.input. We can now
select the MeshDecimate node and adjust parameters to fit our needs, for example, by setting a maximum vertex count
to 100,000. To start the computation, either press the main “Start” button, or right-click on a specific node and select
“Compute”.
Create a MeshDecimate node, connect it, adjust parameters and start computation
By default, the graph will become read-only as soon as a computation is started in order to avoid any modification that
would compromise the planned processes.
Each node that produces 3D media (point cloud or mesh) can be visualized in the 3D viewer by simply double-clicking
on it. Let’s do that once the MeshDecimate node has been computed.
• Double-Click on a node to visualize it in the 3D viewer. If the result is not yet computed, it will automatically
be loaded once it’s available.
• Ctrl+Click the visibility toggle of a media to display only this media alternative from Graph Editor:
Ctrl+DoubleClick on a node
•
Making a variation of the original, high-poly mesh is only the first step to creating a tailored 3D model. Now, let’s see
how we can re-texture this geometry.
Let’s head back to the Graph Editor and do the following operations:
• Right Click on the Texturing node > Duplicate
• Right Click on the connection MeshFiltering.output Texturing2.inputMesh > Remove
• Create a connection from MeshDecimate.output to Texturing2.inputMesh
By doing so, we set up a texturing process that will use the result of the decimation as input geometry. We can now
adjust the Texturing parameters if needed, and start the computation.
The MVS consists of creating depth maps for each camera, merging them together and using this huge amount of
information to create a surface. The generation of those depth maps is, at the moment, the most computation intensive
part of the pipeline and requires a CUDA enabled GPU. We will now explain how to generate a quick and rough mesh
directly from the SfM output, in order to get a fast preview of the 3D model. To do that we will use the nodal system
once again.
Let’s go back to the default pipeline and do the following operations:
• Right Click
on DepthMap >
Duplicate Nodes from Here
(“
>>
” icon) to create a branch in the graph and keep the previous result available.
– alternative: Alt + Click on the node
• Select and remove (Right Click > Remove Node or Del) DepthMap and DepthMapFilter
• Connect PrepareDenseScene.input Meshing.input
• Connect PrepareDenseScene.output Texturing.inputImages
We will now sum up by explaining how what we have learnt so far can be used to work iteratively and get the best
results out of your datasets.
1. Computing and analyzing Structure-from-Motion first
This is the best way to check if the reconstruction is likely to be successful before starting the rest of the process (Right
click > Compute on the StructureFromMotion node). The number of reconstructed cameras and the aspect/density of
the sparse point cloud are good indicators for that. Several strategies can help improve results at this early stage of the
pipeline:
• Extract more key points from input images by setting “Describer Preset” to “high” on FeatureExtraction node
(or even “ultra” for small datasets).
• Extract multiple types of key points by checking “akaze” in “Describer Type” on FeatureExtraction, Feature-
Matching and StructureFromMotion nodes.
2. Using draft meshing from SfM to adjust parameters
Meshing the SfM output can also help to configure the parameters of the standard meshing process, by providing a
fast preview of the dense reconstruction. Let’s look at this example:
With the default parameters, we can preview from Meshing2 that the reconstructed area includes some parts of the
environment that we don’t really want. By increasing the “Min Observations Angle For SfM Space Estimation”
parameter, we are excluding points that are not supported by a strong angle constraint (Meshing3). This results in a
narrower area without background elements at the end of the process (Meshing4 vs default Meshing).
\3. Experiment with parameters, create variants and compare results
One of the main advantages of the nodal system is the ability to create variations in the pipeline and compare them.
Instead of changing a parameter on a node that has already been computed and invalidate it, we can duplicate it (or the
whole branch), work on this copy and compare the variations to keep the best version.
In addition to what we have already covered in this tutorial, the most useful parameters to drive precision and perfor-
mance for each step are detailed on the Meshroom Wiki.
Meshroom does not yet provide an export tool to Sketchfab, but results are all in standard file formats and can easily
be uploaded using the Sketchfab web interface. Our workflow mainly consists of these steps:
• Decimate the mesh within Meshroom to reduce the number of polygons
• Clean up this mesh in an external software, if required (to remove background elements for example)
• Retexture the cleaned up mesh
• Upload model and textures to Sketchfab
You can see some 3D scans from the community here and on our Sketchfab page.
Don’t forget to tag your models with “alicevision” and “meshroom” if you want us to see your work!
Capturing
If this is the first time you are using photogrammetry software, read the following chapter on how to take good photos
for your project.
18.1 Basics
75
Meshroom, Release 0.1
18.2 Details
18.3 Tutorials
More
19.1.1 Meshlab
You can drag and drop different OBJ and PLY files as layers.
So in this case I have a layer for both the final mesh and the SFM points/cameras. Sometimes the mesh smoothing
step can be a little too aggressive so I find it useful to compare between the original mesh and the smooth mesh. If the
77
Meshroom, Release 0.1
mesh looks broken, the PLY sfm data and the OBJ meshes are great for tracing through the pipeline.
clean up / delete / smooth
The first thing you want to do is to rotate your model and align it with the coordinate system.
You can import the obj into Meshlab then go to Filters > Normals, Curvatures ** and **Orientation > Transform:
Rotate ** ** and align it yourself from there. ** **
There might be some parts of the model or the scene you want to remove.
You can select . . . .. then remove. . .
http://www.banterle.com/francesco/courses/2017/be_3drec/slides/Meshlab.pdf
http://
www.scanner.imagefact.de/tut/meshlabTut.pdf
Smooth mesh
If you don’t like the smoothing results from Meshroom, you can smooth the mesh yourself.
http://www.cs.cmu.edu/~reconstruction/advanced.html#meshlab
Tutorials by Mister P. MeshLab Tutorials MeshLab Basics: Navigation
MeshLab Basics: Selection, part one
MeshLab Basics: Selection, part two
Cleaning: Triangles and Vertices Removal
Cleaning: Basic filters
Mesh Processing: Decimation Meshlab Processing: Smoothing
MeshLab Basics: Scale to real measures
19.1.2 Blender
For detailed instructions visit the blender homepage or the blender youtube channel .
Here is a quick tutorial on how to optimize photogrammetry objects inside Blender: How to 3D Photoscan Easy and
Free!
https://www.youtube.com/watch?v=k4NTf0hMjtY
meshing filtering 10:18 / 13:17 blender import
https://www.youtube.com/watch?v=RmMDFydHeso
19.1.4 BlenderLandscape
Addon for Blender 2.79b. 3DSurvey Collection of tools to improve the work-flow of a 3D survey (terrestrial or UAV
photogrammetry). Import multiple objs at once (with correct orientation), for instance a bunch of models made in
Meshroom. https://github.com/zalmoxes-laran/BlenderLandscape
https://github.com/wjakob/instant-meshes
includes quick intro
why do we want to use it? It is a really fast auto-retopology solution and helps you create more accurate meshes
19.1.6 CloudCompare
assets subfolder.
Open Unity and wait for the auto-import to complete.
You might want to optimize your mesh and texture for ingame use.
MeshroomMaya (v0.4.2) is a Maya plugin that enables to model 3D objects from images.
https://github.com/alicevision/MeshroomMaya
This plugin is not available at the moment.
Use the Export to Maya node instead.
Alembic bridge
Import in Nuke/Mari
In menu “NukeMVG > Import Alembic” , .abc file can be loaded. The tool create the graph of camera projection.
Result can be export to Mari via Nuke <-> Mari bridge.
https://pointscene.com/
https://www.pointbox.xyz/
and more. . .
https://groups.google.com/forum/#!topic/alicevision/RCWKoevn0yo
Remote control your camera via USB cable. For use with a turntable and/or Live Reconstruction.
Some manufacturers (Sony, Panasonic, FUJIFILM, Hasselblad. Canon EOS..) provide a free tool for your software
others sell them (Nikon, Canon). Some commercial third party solutions are out there, too.
This list only contains free open-source projects.
1 DigiCamControl (Windows)
• Multiple camera support
http://digicamcontrol.com/download
Supports many Nikon, Canon, Sony SLR models and a few other cameras.
Full list here: http://digicamcontrol.com/cameras
2 Entangle Photo (Linux)
https://entangle-photo.org/
Nikon or Canon DSLRs camera supporting ‘ <http://www.gphoto.org/doc/remote/>‘_ remote capture in libgphoto2
will work with Entangle.
3 GPhoto (Linux)
http://www.gphoto.org/
4 Sofortbildapp (OSX)
http://www.sofortbildapp.com/
5 PkTriggerCord (Windows, Linux, Android)
for Pentax cameras
http://pktriggercord.melda.info/
https://github.com/asalamon74/pktriggercord/
4 Darktable (Windows, Linux, OSX)
http://www.darktable.org/
https://www.darktable.org/usermanual/en/tethering_chapter.html
WifiRemoteControl
For some cameras wifi control can be used.
LMaster https://github.com/Rambalac/GMaster for some Lumix cameras for example.
..image:: ofxMVG.jpg
19.5.1 ofxMVG
19.5.2 CCTag
19.5.3 PopSIFT
Solution: try to reduce the value of maxPoints on the Meshing node to avoid using too much RAM & SWAP
#243 #303
You can speed up the Depth Map process. Here is what you need to do:
Augment the downscale factor to directly reduce the precision.
Reduce the number of T cameras (sgmMaxTCams, refineMaxTCams) will directly reduce the computation time lin-
early, so if you change from 10 to 5 you will get a 2x speedup.
A minimum value of 3 is necessary, 4 already gives decent results in many cases if the density of your acquisition
process regular enough.
The default value is necessary in large scale environment where it is difficult to have 4 images that cover the same
area.(#228)
As of version 2019.1.0 of meshroom, it is possible to do a reconstruction without using the Depthmap node (depthmap
requires CUDA). It is much faster than depth map but the resulting mesh is low quality, so it is still recommended that
the depthmap is used to generate the mesh if possible. This can be done using the following node configuration:
87
Meshroom, Release 0.1
You should use the HIGH preset on the FeatureExtraction node to get enough density for the Meshing. Reconstruction-
parameters
Unexpected exit of Meshroom while processing can cause the “Graph is being computed externally” problem.‘#249‘_
The Start and Stop buttons are greyed out.
Background: When Meshroom is terminated unexpectedly, files are left in the cache folders. When you open such a
project, Meshroom will think, based on the residual files, that parts of the pipeline are computed externally. (This fea-
ture ([Renderfarm](https://github.com/alicevision/meshroom/wiki/Large-scale-dataset)) is not included in the binary
Release 2019.1.0) So the buttons are greyed out because Meshroom is waiting for an external source to compute the
graph. Obviously, this won´t go anywhere. This behaviour can also occur, when you modify nodes in the advanced
mode while the graph is being computed.
To fix this problem, first try to ‘Clear Submitted Status’ by clicking on the bad node (right click->delete data).
If this does not work, also clear the submitted statuses of the following nodes (right click->delete data >>>)
You have a menu on the top-right of the graph widget with “Clear Pending Status” to do it on all nodes at once.
Alternatively, go to the cache folder of your project and delete the contents of the node folders starting with the node
where Meshroom stopped working (marked in dark green). You can keep successful computed results (light green).
Now you can continue computing the graph on your computer.
The import module from AliceVision has problems parsing corrupted image files. Some mobile phone cameras and
action cams/small cameras like the CGO3+ from Yuneec produce images which are not valid. Most image viewers
and editing software can handle minor inconsistencies.
Use tools like Bad Peggy to check for errors in your image files.
e.g. “. . . extraneous bytes before marker 0xdb”.
or “Truncated File - Missing EOI marker” on a raspberry camera
To fix this problem, you need to bulk convert your dataset (this is why downscaling worked too). You can use Irfran-
View File->Batch Conversion or Imagemagick. Make sure you set the quality to 100%. Now you can add the
images to Meshroom (assuming the camera is in the sensor db).
–
drag and drop of images does not work (#149) mouse over the with any photos the cursor is disabled and dropping
photos into the viewport has no effect. Do you run Meshroom as admin? If yes, that’s the cause. Windows disables
drag and drop on applications being run as admin.
–
Note: avoid special characters/non-ASCII characters in Meshroom and images file paths (#209)
Can I use Meshroom on large datasets with more than 1000 images?
Yes, the pipeline performance scales almost linearly. We recommend adjusting the SfM parameters to be a bit more
strict, as you know that you have a good density / good connections between images. There are 2 global thresh-
olds on the Meshing node (maxInputPoints and maxPoints) that may need to be adjusted depending on the
density/quality you need and the amount of RAM available on the computer you use.
Can I use Meshroom on renderfarm?
Meshroom has been designed to be used on renderfarm. It should be quite straightforward to create a new submitter,
see the available submitters as examples. Contact us if you need more information to use it with a new renderfarm
system.
If you shoot a static dataset with a moving rig of cameras (cameras rigidly fixed together with shutter synchronization),
you can declare this constraint to the reconstruction algorithm.
Currently, there is no solution to declare this constraint directly within the Meshroom UI, but you can use the following
file naming convention:
|---- DSC_0002.JPG
|-+ 1/ # sub-folder with the index of the camera
|---- DSC_0001.JPG
|---- DSC_0002.JPG
All images with the same name in different “rig/cameraIndex” folder will be declared linked together by the same
transformation. So in this example, the relative pose between the 2 “DSC_0001.JPG” images from the camera 0 and
camera 1 will be the same than between the 2 “DSC_0002.JPG” images.
When you drop your images into Meshroom, this constraint will be recognized and you will be able to see it in the
CameraInit node (see Rig and Rig Sub-Pose of the Viewpoints parameter).
[error] This program needs a CUDA-Enabled GPU (with at least compute capability 2.0), but Meshroom is running
on a computer with an NVIDIA GPU.
Solution: update/reinstall your drivers Details: #182 #197 #203
The depth map computation is implemented with CUDA and requires an NVIDIA GPU.
#218 #260
[Request] Remove CUDA dependency alicevision/#439
Currently, we have neither the interest nor the resources to do another implementation of the CUDA code
to another GPU framework. If someone is willing to make this contribution, we will support and help for
integration.*
Yes, but you must use Draft Meshing to complete the reconstruction.
Check https://developer.nvidia.com/cuda-gpus
The default parameters are optimal for most datasets. Also, many parameters are exposed for research & development
purposes and are not useful for users. A subset of them can be useful for advanced users to improve the quality on
specific datasets.
The first thing is to verify the number of reconstructed cameras from your input images. If a significant number are
not reconstructed, you should focus on the options of the sparse reconstruction.
1. FeatureExtraction: Change DescriberPreset from Normal to High If your dataset is not big
(<300 images), you can use High preset. It will take more time for the StuctureFromMotion node but it
may help to recover more cameras. If you have really few images (like <50 images), you can also try Ultra
which may improve or decrease the quality depending on the image content.
2. FeatureMatching: Enable Guided Matching This option enables a second stage in the matching pro-
cedure. After matching descriptor (with a global distance ratio test) and first geometric filtering, we retrieve
a geometric transformation. The guided-matching use this geometric information to perform the descriptors
matching a second time but with a new constraint to limit the search. This geometry-aware approach prevents
early rejection and improves the number of matches in particular with repetitive structures. If you really struggle
to find matches it could be beneficial to use BRUTE_FORE_L2 matching, but this is not good in most cases as
it is very inefficient.
3. Enable AKAZE as DescriberTypes on FeatureExtraction, FeatureMatching and
StructureFromMotion nodes It may improve especially on some surfaces (like skin for instance).
It is also more affine invariant than SIFT and can help to recover connections when you have not enough images
in the input.
4. To improve the robustness of the initial image pair selection/initial reconstruction, you can use a SfM with
minInputTrackLength set to 3 or 4 to keep only the most robust matches (and improve the ratio in-
liers/outliers). Then, you can chain another SfM with the standard parameters, so the second one will try again
to localize the cameras not found by the first one but with different parameters. This is useful if you have only
a few cameras reconstructed within a large dataset.
1. DepthMap
You can adjust the Downscale parameter to drive precision/computation time. If the resolution of your
images is not too high, you can set it to 1 to increase precision, but be careful, the calculation will be ~4x
longer. On the contrary, setting it to a higher value will decrease precision but boost computation.
You can choose to use one or multiple describer types. If you use multiple types, they will be combined together
to help get results in challenging conditions. The values should always be the same between FeatureExtraction,
FeatureMatching and StructureFromMotion. The only case, you will end up with different values is for testing and
comparing results: in that case you will enable all options you want to test on the FeatureExtraction and then use a
subset of them in Matching and SfM.
StructureFromMotion may fail when there is not enough features extracted from the image dataset (weakly
textured dataset like indoor environment). In this case, you can try to augment the amount of features:
• DescriberPreset to High or Ultra in FeatureExtraction
• Add AKAZE as DescriberType on FeatureExtraction, FeatureMatching and
StructureFromMotion nodes
Using more features will reduce performances on large datasets. Another problem is that adding too much features
(less reliable) may also reduce the amount of matches by creating more ambiguities and conflicts during features
matching.
• Guided Matching parameter on FeatureMatching is useful to reduce conflicts during feature matching
but is costly in performance. So it is very useful when you have few images (like a cameras rig from a scan
studio).
Meshroom supports most image formats, including many RAW formats such as ‘.exr’, ‘.rw2’, ‘.cr2’, ‘.nef’, ‘.arw’,. . .
The image importer is based on OpenImageIO, so all formats supported by OpenImageIO can be imported to Mesh-
room. However it is recommended to use ‘.jpg’, ‘.jpeg’, ‘.tif’, ‘.tiff’ or ‘.png’ at the moment.
Note: On some datasets the reconstruction quality could be reduced or cause unexpected interruption of the pipeline.
(#G) Convert your RAW image to ‘.jpg’, ‘.jpeg’, ‘.tif’, ‘.tiff’ or ‘.png’ to resolve this problem.
It is possible to reproject textures after re-topology and custom unwrap. The only constraint is to NOT modify
scale/orientation of the model, in order to stay in the same 3D space as the original reconstruction.
To retexture a user mesh, your need to remove the input connection on Texturing node’s inputMesh (right click
connection > Remove) and write the path to your mesh in the attribute editor. If you have custom UVs, they will be
taken into account.
You can also duplicate the original Texturing node (right click > Du-
plicate) and make changes on this copy. It should look like this:
(optional) You can also set ‘‘Padding‘‘ to 0 and check ‘‘Fill Holes‘‘ instead if you want to completely fill texture’s
blank space with plausible values.
20.13 Troubleshooting
References
Text publications
...
Videos
Meshroom live reconstruction (LADIO project)
https://www.youtube.com/watch?v=DazLfZXU_Sk
Meshroom: Open Source 3D Reconstruction Software
https://www.youtube.com/watch?v=v_O6tYKQEBA
How to 3D Photoscan Easy and Free!
mesh filtering 10:18 / 13:17 blender import
https://www.youtube.com/watch?v=k4NTf0hMjtY
Meshroom: 3D Models from Photos using this Free Open Source Photogrammetry Software
https://www.youtube.com/watch?v=R0PDCp0QF1o
Free Photogrammetry: Meshroom
https://www.youtube.com/watch?v=NdpR6k-6SHs
MeshRoom Vs Reality Capture with blender
https://www.youtube.com/watch?v=voNKSkuP-RY
MeshRoom and Blender walkthrough
https://www.youtube.com/watch?v=VjBMfVC5DSA
Meshroom and Blender photoscanning tutorial (+ falling leaf animation)
https://www.youtube.com/watch?v=3L_9mf2s2lw
Meshroom Introductory Project Tutorial
https://www.youtube.com/watch?v=bYzi5xYlYPU
95
Meshroom, Release 0.1
Glossary
97
Meshroom, Release 0.1
About
Meshroom is a free, open-source 3D Reconstruction Software based on the AliceVision framework. AliceVision is a
Photogrammetric Computer Vision Framework which provides 3D Reconstruction and Camera Tracking algorithms.
AliceVision aims to provide strong software basis with state-of-the-art computer vision algorithms that can be tested,
analyzed and reused. The project is a result of collaboration between academia and industry to provide cutting-edge
algorithms with the robustness and the quality required for production usage.
In 2010, the IMAGINE research team (a joint research group between Ecole des Ponts ParisTech and Centre Scien-
tifique et Technique du Batiment) and Mikros Image started a partnership around Pierre Moulon’s thesis, supervised
by Renaud Marlet and Pascal Monasse on the academic side and Benoit Maujean on the industrial side. In 2013, they
released an open source SfM pipeline, called openMVG (“Multiple View Geometry”), to provide the basis of a better
solution for the creation of visual effects matte-paintings.
In 2009, the CMP research team from CTU started Michal Jancosek’s PhD thesis supervised by Tomas Pajdla. They
released Windows binaries of their MVS pipeline, called CMPMVS, in 2012.
In 2009, INPT, INRIA and Duran Duboi started a French ANR project to create a model based Camera Tracking
solution based on natural features and a new marker design called CCTag.
In 2015, Simula, INPT and Mikros Image joined their efforts in the EU project POPART to create a Previz system.
In 2017, CTU joined the team in the EU project LADIO to create a central hub with structured access to all data
generated on set.
23.1.2 Partners
99
Meshroom, Release 0.1
We build a fully integrated software for 3D reconstruction, photo modelling and camera tracking. We aim to provide
a strong software basis with state-of-the-art computer vision algorithms that can be tested, analyzed and reused. Links
between academia and industry is a requirement to provide cutting-edge algorithms with the robustness and the quality
required all along the visual effects and shooting process. This open approach enables both us and other users to
achieve a high degree of integration and easy customization for any studio pipeline.
Beyond our project objectives, open source is a way of life. We love to exchange ideas, improve ourselves while
making improvements for other people and discover new collaboration opportunities to expand everybody’s horizon.
This manual is a compilation of the resources found on alicevision.github.io, breadcrumbs of information collected
from github issues, other web resources and new content, created for this manual.
WORK IN PROGRESS! (last update 03.06.19)
You want to help? Missing something?
You are welcome to comment and contribute. This document is in “Suggest edits” mode.
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License. This is a Mesh-
room community project.
All product names, logos, and brands are property of their respective owners. All company, product and service names
used in this document are for identification purposes only. Use of these names, logos, and brands does not imply
endorsement.
23.3 Acknowledgements
A big thanks to the many researchers, who made their work available online so we can provide free, additional back-
ground information with this guide through references.
And finally thank you for using Meshroom, testing, reporting issues and sharing your knowledge.
To all Meshroom contributors: keep up the good work!
23.4 Contact us
23.5 Contributing
Alice Vision relies on a friendly and community-driven effort to create an open source photogrammetry solution.
The project strives to provide a pleasant environment for everybody and tries to be as non-hierarchical as possible.
Every contributor is considered as a member of the team, regardless if they are a newcomer or a long time member.
Nobody has special rights or prerogatives. The contribution workflow relies on Github Pull Request . We recommend
to discuss new features before starting the development, to ensure that development is efficient for everybody and
minimize the review burden.
In order to foster a friendly and cooperative atmosphere where technical collaboration can flourish, we expect all
members of the community to be courteous, polite and respectful in their treatment of others helpful and constructive
in suggestions and criticism stay on topic for the communication medium that is being used be tolerant of differences
in opinion and mistakes that inevitably get made by everyone.
Join us on Github
https://github.com/alicevision/
23.7 Licenses
• Python https://www.python.org Copyright (c) 2001-2018 Python Software Foundation Distributed under the
PSFL V2 .
• Qt/PySide2 https://www.qt.io Copyright (C) 2018 The Qt Company Ltd and other contributors. Distributed
under the LGPL V3 .
• qmlAlembic https://github.com/alicevision/qmlAlembic Copyright (c) 2018 AliceVision contributors. Dis-
tributed under the MPL2 license .
• QtOIIO https://github.com/alicevision/QtOIIO Copyright (c) 2018 AliceVision contributors. Distributed under
the MPL2 license .
Documentation
Meshroom is a free, open-source 3D Reconstruction Software based on the AliceVision framework.
AliceVision is a Photogrammetric Computer Vision Framework which provides 3D Reconstruction and Camera Track-
ing algorithms. AliceVision aims to provide strong software basis with state-of-the-art computer vision algorithms that
can be tested, analyzed and reused. The project is a result of collaboration between academia and industry to provide
cutting-edge algorithms with the robustness and the quality required for production usage.
A
Alicevision, 97
S
SIFT, 97
103