3-D Photography

How to use photographs and Photosynth to create 3-D mesh models

      Table of Contents





Active Project Notes


Ever find an interesting object and want to show it to your friends and colleagues? Pictures might capture some aspects of it but can leave certain details out. It’s especially hard to capture the depth and dimensionality of an object in a single photo. You could take dozens of photos, but something about looking at a slideshow of one object isn’t appealing. Photosynth, ARC3D, and Toolkit allow you to display these photographs in an intuitive, 3-dimensional space.

This document will show you how to take a collection of photographs and transform them into a 3-D model. This model can be viewed on a standard computer with open-source software or be displayed in 3-D virtual environments such as the Duke immersive Virtual Environment (DiVE). You can also recreate a replica of the object with a 3D printer.

The general computing requirements consist of a Windows compatible device. A Windows

virtual machine run on a Mac is acceptable. The programs are all free and available online.

Workflow Overview:

Photos-->Pointcloud-->Mesh-->Texture-->3D Print/Virtual Reality

Photogrammetry Forum:

For all questions and for a lot more information



The directions below show, step-by-step, how to reconstruct an object in 3-dimensions from a set of photographs. Steps 1-8 detail the simplest process available which uses Photosynth to create the point cloud and Meshlab to import it and create the mesh. There are several more advanced programs available that you can find out more about on the pgrammetry forum site linked above. I recommend photosynth toolkit by Henri Astre and have included basic instructions on using it. Once you get a handle of steps 1-8 below, start exploring other options in order to learn how to create a highly detailed object.

1. Find an object

2. Photograph it!

        a. Take 1-300 photographs of the object for Photosynth or 5-10 for ARC3D

        b. Take the photographs along the perimeter of the object with the object as the center

        c. Zoom in and focus on areas of interest

        d. Overlap (and parralex) is key!

        e. Don’t turn the camera sideways-some programs can’t figure this out even if you try to manually rotate the image back

Alternatively, Video it!

a. Take an HD video of the object

b. Convert movie to images

a. Quicktime Pro 7: Export-->Image Sequence-->Choose # of frames

b. Virtual Dub? Have not attempted yet

3. Photosynth

        a. Create a Microsoft live account (free)

        b. Load photos onto Photosynth Server

        c. Explore point cloud

                i. The better the point cloud, the better the mesh

                ii. Experiment to figure out what makes better point clouds

4. Meshlab (or viewing software of choice)

1. Import pointcloud from photosynth

        Filters-->Create New Mesh Layer-->Import Photosynth data

        Input URL of your photosynth

        Deselect download images (unless you want them, then specify image directory)

        Deselect show cameras

        Click Points Icon

        Render-->Color-->Per Vertex

2. Remove unwanted points

Point select>Delete Point

Hold Control or Command to select multiple groups

Filters--->Selection--->Select Faces by Color. You can select using                 RGB and HSV, and tweak the range of values.

3. Save/export cleaned up image as .ply

4. Compute Normals [skip this if using photosynth toolkit (4_C, below):

Filters>Normals, Curvatures, and Orientation>Compute Normals for Point sets

This filter has a single parameter for the number of neighbors.

Default 10 neighbors but use 100.

5. Generate a mesh using the Poission reconstruction. The function can be found under: Filters>Point Set>Surface Reconstruction: Poisson Reconstruction.

This filter has four parameters. Click the help button for more information on the parameters. It is important to read about their function as altering the parameters profoundly changes the results. I typically use the following values: Octree Depth = somewhere between 10-12, Solver Divide = either 7 or 8, Samples per Node = 1 (default), Surface offseting = 1 (default). This can be quite glitchy-I’ve found that Octree Depth=11 and Solver Divide=7 tends to work so I’ve stuck with that.

6. If the resulting mesh is a bubble (which it very often is), remove the unwanted parts of the bubble. This task is performed using the triangle selection tool and the triangle deletion tool. Both of these are on the tool bar. They can be found near the point selection and deletion tools.

7. Remove extra edges with faces. This function can be fund under Filters> Cleaning

and Repairing> Remove edges with faces longer than...

8. Color:

Filter>Sampling>Vertex Attribute Transfer Function

Make Sure the original pointcloud coordinate system is selected up top and the poisson mesh is selected down below

Render-->Color-->Per Vertex

Click light bulb

Other programs: (notes constantly being updated)


Can simply transfer color attribute from point cloud to mesh or can ‘paint’ on a texture from the original photos. Using the original photos allows you to paint on the tiny details onto the 3D mesh you just generated but is also more time and not “true” 3D reconstruction.

Vertex attribute transfer:

1. Meshlab (pointcloud to mesh)


1. Blender (free)

2. Maya (most familiar)

3. Autodesk Max3D

        Photosynth toolkit 4 is geared for this

can get maya and autodesk free as a student

Photosynth Toolkit

        For Dense point clouds        http://www.visual-experiments.com/2010/08/22/dense-point-cloud-created-with-photosyth-and-pmvs2/

A how-to video for photokit: http://blog.neonascent.net/archives/photosynth-toolkit/

Does it have to be 100% synthy?

How to run PhotoSynthToolkit4 (Windows OS 64 bit version)

Required software:



Meshlab (optional)

1) Create a PhotoSynth synth from a set of photos.

2) Download PhotoSynthToolkit4:


main site:


3) Decompress PhotoSynthToolkit4.zip to C:\


4) Run script: 1 - Download Synth.wsf

4.1) Enter your Photosynth URL

4.2) Choose output path:


5) Copy all the original photos (NOT the downloaded thumbs), used to create your synth to:


6) Run script: 3 - Prepare for PMVS2.wsf

6.1) Choose your input path:


6.2) The program will ask to choose some point cloud creation options*, and finally will create a file named: launch_pmvs.bat

7) Run launch_pmvs.bat

The program will take some time to run. When it ends you will find a PLY file in:


How to run PhotoSynthToolkit4 (Windows OS 32 bit version)

Go to: C:\PhotoSynthToolkit4\bin\PMVS2

erase and replace the 64 bit files:



for their 32 bit version:



You can find this last two files inside this zip:


main site:


*Notes on pmvs_options.txt

Josh Harle’s Video showing results of different parameters: http://vimeo.com/15223228

Level: 0 will give you a denser point cloud but will be slower than using the default 1 option

wsize: 7 default, but go up to 9 if possible


Using masks allows you to inform the software which part of your image you find interesting and to limit the noise coming from irrelevant points. This not only produces a point cloud that is semi-cleaned up, it also can increase the density for the region of interest.

Another good, quick tutorial from Josh Harle: http://vimeo.com/18517975

Upload-only websites

        There are several websites that allow you to simply upload your photos and they will send you a point cloud. The quality is again dependent upon your photographs as well as how each site will compute them. There are both free and pay alternatives out there and I’ve included a brief description of some of the free ones below.

Positives: Free, makes point cloud for you (one less step), can do on mac os?

My 3D Scanner


Can take videos! (not that complicated since they just framegrab but still, makes it quick)

Zip and upload

Wow worked pretty well

        +Very dense point cloud, probably ½ of photosynth toolkit at 0 dense level

        +they auto clean it up, very, very well

        - do they own my photos? i don’t understand the TOS

        - only small photosets?

Post on pgrammetry forums:

300mb .zip file limit

10 mb per photo

+they give feedback

some of my results in their tutorials?

Get 180 degree FOV at most so need to break 360 object into multiple faces then combine:


CMP SfM Web Service


Need to email to ask for username

Zip and upload

Less documentation/turials but say they’re working on it.

Am attempting a few trials and will report back


        5-10 photographs taken in sequence around a circumfrence

        create account (free)



        Download .zip, unzip

4_B. Meshlab        

Click on .v3D


Select photographs on right side that have good depth maps (red)

Put in high subsample (11?)

Click Fast merge?

Select all images on the right (or those that are good according to heat map?)

Click export to .ply

Then click ok

Very glitchy-select high subsample and low resolution and fast-merge

3_D. Bundler Noah-Snavely options?

3_E. Photofly


Worked pretty well...

needs small amount of photos

Is there a way to get output into meshlab?

3_F. Insight 3d


Haven’t tried yet. Open-source but looks like it crashes often.





$$$$ ($1,500-2,500)

Demo only allows you to use photos from company

Supposedly can contact local reseller who will give you code that enables you to use your own photos for the demo.

Aeroscan (aeroscan.com)

Goldingart has a pretty amazing reconstruction from it (in example gallery)

Free to upload and reconstruct but costs money to export into formats you want

~$100 per download

No portrain photos, only landscape. Keep photo size and zoom level constant.

6. DiVE

        a. Convert .ply to Maya compatible extension

        b. Import into Virtools

        c. Provide figure dimensions

        d. View in DiVE

6_B. 3-D printing




Great link with video tutorials


A Meshlab team blog describing their attempts at photosynth pointcloud-->mesh experience.



A good general meshlab tutorial?

ARC3D-->Meshlab Links:


Quick overview of process


More detailed look at importer


A look at some of the parameters


Noah Snavely’s Bundler version of photosynth





ESPNs 3D scans of student athletes bodies:





Photosynth-->DiVE Models

Statue Photosynth: http://photosynth.net/view.aspx?cid=de9d2943-5628-4fd5-aba3-c0ce3b6eaf4d




ARC3D (only 8 photos!)

ARC3D, Duke Chapel (20 photos)


Toolkit4, Duke Chapel 20 photos, 3.5MB High-Resolution

(captured much more detail of the front face vs. the above ARC3D reconstruction but did not capture the entire structure as well)

Toolkit2 (Thumbnail, low resolution version so less details but very photorealistic color-wise)

Another Toolkit rendering, look familiar? (low resolution thumbnails)

Photofly. Photo Editor. Statue with 8? photos (less than arc3d). Good sides...best art in back

Active Project Notes



Video side by side. Video walk around on left and screen capture of similar path of model. Screen shot them side-by-side.

David McMullen

UMDNJ Robert Wood Johnson Medical School

3-D Photography

Regular digital photographs can now be used to create models for use in virtual environments (such as the DiVE) and 3-D printers. This talk will demonstrate a simplified process for transforming a series of object-oriented photographs into 3-D models. Widespread applications include art and archaeology to businesses interested in a quick way to share a prototype. A tutorial will follow the talk for those interested in hands on help.

3D Printing


email author and let him know about photogrammetry (http://topics.nytimes.com/top/reference/timestopics/people/v/ashlee_vance/index.html?inline=nyt-per)

Work on resizing in blender, then see what shapeways says

Online Computing


can get up to 64gb RAM but need to try it out first


3D video????


Microsoft Kinect:



can use 2 kinects to get a ~180 degree 3d video model

Video Game Animation


Pretty cool use of 360degrees of video cameras to recreate facial expressions for a video game

3D Photo Notes:

Get plenty of background in photos b/c it helps to provide context for images to find relationships. Be sure to delete these along the way when doing the reconstruction though since they’re not that useful and cause huge bubbles. Make sure the background isn’t moving or just huge (out a window) since that isn’t that useful.


Take HD video, walk around object. Extract stills? would that work well enough???

Links to try

Lana’s Face


by bhowiebkr under Creative Commons



Goutz (Henri’s)




Good b/c will see if gaps are filled in or not

Me: http://photosynth.net/view.aspx?cid=c37895da-1102-4b04-8078-6fd9ac327f95

Statue: http://photosynth.net/view.aspx?cid=7d6b94ba-7fcb-4acf-8073-7b84b14151c4

Brain5: http://photosynth.net/view.aspx?cid=bbe2f869-8422-424a-80c2-ac78363cfe68

Brain 4: http://photosynth.net/view.aspx?cid=76a31734-ad43-4208-9349-14cca0f77633

Statue of Liberty:

Jason’s mesh one:




good point cloud, a statue


Notre Dame:

alright point cloud, a good outdoor structure


Any of M4’s synths







La Pieta:




Synth? http://photosynth.net/view.aspx?cid=6d259a2c-9096-4a6a-a8c0-9dae655171ec