When performing detailed comparisons of two objects using photographs, ideally they should be from same viewpoint (angle and distance). Otherwise, perspective distortion can come into play making even the same object look quite different from different distances.

While researching phtogrammetry, I ran across VisualSFM (SFM=Structure From Motion). It's a free program than can use a set of pictures taken at various distances and angles and calculate camera viewpoints and a basic 3D object model, or at least a "3D point cloud".

From that 3D model, measurements could be made and compared against another 3D model, effectively factoring out the uncertainties of 2D image comparisons (in theory at least).

Here's a screen capture of the 3D Dense Reconstruction of a 30-picture set of my Darth Ugly ANH helmet:



and a 2-minute slide-based overview (no sound) of VisualSfM and my using it to create the above 3D model:

https://drive.google.com/file/d/0B13YyeojRHKeNlZUUkNqVGpwb3M/edit?usp=sharing

I've done some testing of the Structure From Motion (SFM) program applied to consecutive frames extracted from a youtube video of the eFx Vader.

Vader can be a tricky subject for this program to process. The program needs to track specific points across multiple images. For a smooth surface, few features will be found. Glossy surfaces create reflections which can confuse the algorithm. It seems to perform better if there is also a background that has trackable features.

What does not work for this algorithm is a fixed camera with a 360 degree turning of the helmet (a moving helmet and a fixed background confuses the tracking algorihtms). What also doesn't work is a fixed helmet and camera images taken from multiple angles but without varying distance. For these cases the program seems to map the tracked features to the surface of a generic sphere in the absence of depth cues:

Ideally a single model will be created from the entire set of input images. But when the algorithm cannot relate the current image to the previously processed ones, there is a break; a new model is then created at that point using only the current and subsequent images, not the previously processed ones. You can scroll through various models created, but having multiple models is not ideal because it's not benefiting from the complete set of images, and only a portion of the helmet will be modeled.

But when a large number of images contribute to a single model with good coverage across the helmet, the results can be pretty good.

After running "SfM > Pairwise Matching > Compute Missing Match" and while running "SfM > Reconstruct Sparse" you can watch the 3D model build as each image is being processed from "View > N-View 3D Points" :

Then after running "SfM > Reconstruct Dense" (which can take a while to run) and selecting "View > Dense 3D Points" you get a better quality 3D rendering:

Here is that model viewed from different angles.

The model can of course be rotated and viewed from any arbitary angle and distance, though the quality of the image is better along the viewpoints that correspond to the input images. For the following, there were few or no images taken along these viewpoints, so the 3D model doesn't look as good from those angles.

While viewing the 3D model, you can change the pixel size on the fly from a coarse square down to a single point:

When using a small pixel size, it's easier to be able to see through the surfaces and effectively give an X-Ray vision perception. For example, from this profile view you can see the slope/curvature inside the nose.

The accuracy of the points in the 3D model would not rival having a caliper measurement of an available helmet. But for a rough rendering of basic proportions, it's not bad. And it's interesting to be able to rotate, zoom, and translate the 3D model from any arbitrary viewpoint.

Have you ever tried using 123d catch for this? They make pretty good 3D scans from pictures. The program is free to download here:

http://www.123dapp.com/catch

I had also looked at 123D Catch but it looks like Autodesk has worked on this a lot since I first tried it. Autodesk has provided a vast assortment of modules and applications to provide an end-to-end 3D capture, design, sharing, and printing capability. Some of it like the social "facebook" sharing angle is fluff to me, but I recognize they are trying to monetize this effort.

I am more interested in this from the technical perspective to sample the state of the art currently in capturing a 3D model from photographs. Both Autodesk 123D Catch and VisualSFM likely use very similar algorithms to do this. VisualSFM has the advantage of being able to tweak the algorithm that does the feature tracking and run it locally (Autodesk requires that you transmit the photographs to their servers where the real number crunching is done). As such, VisualSFM has the advantage of seeing the algorithms working live.

Autodesk does seem to have a better capability of retaining the original colors and surface textures. Both however don't do a great job of capturing the lines and curves of the geometry of the Vader helmet. The dome especially given its lack of features and black/glossy surface can't be captured by either. It's like trying to extract features from the 2001 A Space Odyssey monolith. This could probably be accommodated by sprinkling powder or glitter (for example) on the surface to give the algorithms something to track and then edit those out later with another tool.

For both it seems to be a matter of trying to find exactly the optimum photographic setup that works with each application.

Here's a couple of quick captures I did of my DPDLX in 3 different setups using Autodesk 123D Catch:

1) Inside - this provided the best overall capture. I believe it picked up what could be some surface features obscured by paint, but also interspersed with artificial artifacts from the algorithm interpolating and "guessing":



2) Outside (direct sunlight) - Interesting in this is that it captured the "nose droop" (the downward angle at the tip of the nose) which is not how it was painted and not immediately visible but would be accurate according to how the original helmets were molded.



3) Outside (indirect diffuse sunlight) - This gave the worst results surprisingly, looking like Vader was just sent through an underpowered Star Trek transporter.

It looks like a new form of art... or that Vader stepped out of a steam bath. :-)

LOL!

Honestly: I'm an expert in 3D modeling and have tried out some of this cheap amateur tools incl. Autodesk123 and Kinect based solutions, but all of them are unuseable for really accurate scanning work and not worth time and effort.

And like any Sith Lord, Vader is able to conceal his true nature from the unscrupulous observer!