Photogrammetry: An approach

This is the work flow I'm using to convert photos into 3d models for rendering in blender.  Obviously this isn't the best or only way to do this but it's what I've been using so far.  This also isn't intended as a full tutorial. 
​
There are 2 tools in use:

• Regard3D - Free open source structure from motion software
• Meshlab - Free open source application for processing triangular meshes

Step1: Take photos

Take lots of photos of your chosen subject.  

• Make sure there is plenty of overlap between the photos so the algorithm can detect and match features in multiple images.  
• This works best on a overcast day.
• Use your cameras sharpest f-stop.  If not sure use f8.
• Use a low iso to get clean sharp image.

 

Step2: Generate mesh using Regard3d

Regard3D is used to convert the image set into a 3d mesh.  This is done in a series of steps.  These are shown in the project window below, but basically you load the image set, generate matching points between images, triangulate the result, then densify it.

Step3: fix up in MeshLab

In MeshLab we run a series of filers over the data to convert it into a model we can load into other applications.

• Screened Poisson Surface Reconstruction - Generate a water tight mesh, set reconstruction depth to 12
• Remove any vertices that don't contribute to the model
• Quadric Edge Collapse Decimation - Set the target number of faces to something reasonable
• Remove Faces from Non Manifold Edges - A edge should only be attached to 2 faces
• Parameterization + texturing from registered rasters  - Set texture size to something big
• Export mesh as .obj file

Step4:Results

The model was imported into Blender and a simple cycles shader node created using the texture.  The data can then be rendered or modified as needed.    I definitely could have taken some time to clean up the edges.

Here's another rock.  Or is it a giant mountain range.  Once in the 3d package, it could be anything :-)