I’m looking for an approach to reconstruct a 3D model of eye glasses (without lenses) in a fast and accurate way with as little as possible manual post-processing (as in no Meshlab if possible!).
One difficulty for laser, structured light or traditional photogrammetric approaches is the diversity of the material (from highly specular metal to very diffuse black plastic), the impossibility to apply powder (could not resell the product afterward), the geometrically thin structures without texture, etc. I also want to capture the texture of the glasses to have a good rendering of “the real thing”.
One aspect of this particular problem that might help is the very constrained geometry/topology of eye glasses. Is this something that could be exploited by some kind of synthesis image-to-CAD algorithm that could use the information contained in a few images to generate a faithful 3D model?
Thanks
Hi,
You can try with the attached paper. The authors are focusing their research in virtual 3D eyeglasses try on system driven by a 2D internet human image. They are able to generate a full 3D model from a 2D glasses image.
Regards.
Article A mixed reality system for virtual glasses try-on
Thanks but if I'm reading correctly the beginning of section 3:
"Given a 3D virtual glasses, the purpose is to fit the 3D glasses into a live video accurately in realtime. Some points are marked offline on 3D models of spectacles; however their fitting on the face is done in real-time in each image of the input sequence"
seems to indicate that they use the 3D models to generate the a 2D projection of what would be seen based on the position of the head which they are tracking.
My main problem is precisely the acquisition and generation of a faithful 3D model (with texture) of the eyeglasses to use in a system such as the one Yuan describes in its paper. Given that my clients will need to generate many hundreds of such model, I need to develop a very fast and almost automatic method for this.
This is a previous paper. They are now generating 3D models from 2D images and then using this try-on virtual system. You may contact them through ResearchGate.
Did yoy took a look to Shape from Silhouettes method.
You use SfM for calibration with a pattern and then background substraction and shape from silhouette.
Hello Pierre,
Thanks for the suggestion. It might work but as the structures are quite thin, I'm not sure how many shots would ne needed. I tried reconstructing with the Agisoft Photoscan software but there was not enough features matching from image-to-image and some were simply wrong because of specular reflexions.
I could still use the software to position the shots based on targets. Do you know a software I could try for the silhouette reconstruction part?
I found this project https://github.com/NewProggie/Restore
Else you can try the commercial product 3DSom. THey claims "Fast precise wireframe generation – retains the 3DSOM patented “shape-from-silhouette” engine, with no voxel grid aliasing." http://www.3dsom.com/feature-list/
Hello Pierre,
I will definitely check these! I work for a small applied research computer vision consulting organism affiliated with a college and given our limited resources, the best solution for us is often to buy the right product of find the right library.
Merci encore,
Bruno
You're welcome, Please provide feedback about your results ;-)
Regarding the library side, if you wanna do SfM or Multiple View Geometry computation => You can take a look to https://github.com/openMVG/openMVG
Just to "close" this post on my side, the 3DSOM software did a great job of reconstructing 3D models of eye glasses with the shape from silhouette method!
- Badges
- Science topic