28
$\begingroup$

Photogrammetry is really frustrating. The good tools are really expensive, and the free tools are either really old, or really difficult to get working because they're pet-projects... or they're Autodesk.

Current workflows use one tool to reconstruct the structure. Often either a Structure From Motion approach, or a multi-view stereo approach. Then another tool is used to turn those results into a mesh, almost always MeshLab. And another tool to texture using the photographs, often MeshLab and/or Blender.

Blender has all the functionality to identify features in an image and match those features across multiple images. We do it with motion tracking all the time. Blender also has the functionality to project those features into 3D, as with the empties that get produced when doing any kind of motion tracking. Blender also has the functionality to turn a point-cloud into a mesh.

Are there any tools/addons for doing photogrammetry or digital reconstruction from within Blender, without external applications?

$\endgroup$
13
  • $\begingroup$ Are you actually asking a question on how to do photogrammetry in Blender or just looking for people who might be? $\endgroup$ Nov 19, 2015 at 17:17
  • $\begingroup$ My current understanding is that Blender does not currently have any features that will construct a mesh from a set of photographs or video. If that's not the case, then yeah I'd like to know how to do that. I'm not talking about how to use photographs to manually model, that's prety well understood. I'm interested in precise photogrammetry within Blender. $\endgroup$
    – Matt
    Nov 19, 2015 at 17:24
  • 2
    $\begingroup$ (My experience) The "beginning of something" can bee seen at the end of this video. He was able to get the point cloud of a series of images (video frames). Next step is meshing the pointcloud. There are some interesting projects around the web, but I haven't tried them yet. Smart Unwrap and Projection paint must be automated but they are almost ready $\endgroup$
    – Carlo
    Nov 19, 2015 at 18:50
  • 4
    $\begingroup$ This isn't a valid question, should be re-worded. $\endgroup$
    – ideasman42
    Nov 20, 2015 at 10:14
  • 1
    $\begingroup$ If I already understood that it was invalid, I wouldn't have written it this way. I.e. I'm going to need more information about what makes it invalid, before I understand how to rewrite it. $\endgroup$
    – Matt
    Nov 20, 2015 at 13:57

3 Answers 3

26
$\begingroup$

Although quite labor intensive, and far from automatic (but certainly cheap), it would be possible imagine a workflow using tracking information to reconstruct objects based on photographs.

Using a rotatory platform photograph an object. Bring the image sequence into the Movieclip editor and track as many of the object's features as you can.

enter image description here

Solve the scene and setup a tracking scene.

On the 3D viewport enable Motion Tracking.

Instead of your object moving you'll be presented with a camera that moves around a bunch of markers (the tracked points on the image sequence).

Using those markers you'll have some guidelines to reconstruct the geometry of your object.

enter image description here

enter image description here

Or you can select all your trackers and convert them to vertices by using 3D Markers to Mesh.

That will create an object with a point cloud of unconnected vertices. You can then connect the dots...

enter image description here

enter image description here

Once you are done with the modeling, you can use some of the same photos to project the textures onto the UV unwrapped model using the UV project modifier and the camera as projector.

$\endgroup$
12
$\begingroup$

This might be no direct answer to your question, but may help to solve your problem.

You stated:

Photogrammetry is really frustrating. The good tools are really expensive, and the free tools are either really old, or really difficult to get working because they're pet-projects... or they're Autodesk.

I do not think that this statement is actually true. I have worked with several open-source Structure from Motion (SfM) tools over the past few years. And they achieve impressive results. Have you used one of the following state-of-the-art open source SfM tools:

I have wrote a Blender Addon that allows you to import photogrammetry results from different libraries including Colmap, Meshroom, MVE, OpenMVG, OpenSfM and VisualSfM:

https://github.com/SBCV/Blender-Addon-Photogrammetry-Importer

Here are a few images showing some imported reconstruction results.

Imported SfM Result Imported SfM Result with animation Imported SfM Result with Mesh

$\endgroup$
2
  • $\begingroup$ I've not used any of these tools. The only ones I could find back in the day were a pair of abandoned projects. I've been using Agisoft Photoscan (now called Metascan) lately, and I'm still disappointed with how difficult it is to get good results. I'll certainly check out your suggestions. $\endgroup$
    – Matt
    Aug 2, 2019 at 14:58
  • $\begingroup$ Can recommend Meshroom and COLMAP as well. Just be aware that COLMAP uses a LOT of hard disk space saving project data / intermediate results. $\endgroup$
    – Robert Gützkow
    Aug 6, 2019 at 19:50
2
$\begingroup$

Just to add. The easiest way to get more dots on an object is to lower the size of the trackers to 5 and then the recognition are to 20. Hit automatic detection then hen change the distance from 120 to 10 and maybe lower the threshold. Unfortunately Blender on 2.78 was a bit unstable so with about 700 markers it would crash. Try to only do 30 frames.

An interesting hack is to project markers onto the object and move around it with the camera instead of rotating it. I managed to capture parts of a white walled room using an AV projector with a dot field image.

It would be really nice if someone would extend the system to allow markers of a different type so that all of the object could be tracked.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .