Autodesk-forge, Autodesk-realitycapture: Very scrambled 3D models!!! are being generated

Question

Why is Autodesk Forge Reality Capture API generating such poor models from these images?

And more importantly, what can be done to get better quality models, going forward?

http://cloud1.tri-di.com/Good_images_yielding_poor_3D_model/photoscene-1xWRQ4B7px17ZMmgcjkqY6uPBtYGZJREH9wAPSPKaIQ/obj_model

http://cloud1.tri-di.com/Good_images_yielding_poor_3D_model

Answer 1

I am not expert of Recap, but this blog post What makes photos good for photogrammetry may answer your question. Your model seems a little transparent, and shiny to do photogrammetry, the limitation is mentioned in the blog as follow:

KNOWN TECHNOLOGICAL LIMITATIONS OF PHOTOGRAMMETRY:

The photogrammetry process does not work on shiny, transparent or highly glossy objects. This is because these objects reflect the surrounding which makes the same point on one photo to be in a different color on another photo, making it impossible for the algorithms to calculate the scene.

Answer 2

Per our Engineering, quite a lot of conditions of this capture are bad for photogrammetry and they are:

• The lighting condition is not constant. ie, from shot to shot the color of the light is wildly different. Also sometime the light source themselves is over exposed.

• No enough feature on both background and subject. Since the subject is just a single color for the most part, it is difficult for the algorithm to detect features. Similarly the mostly white background offer no feature to help guide the algorithm

• Added markers are not unique. While it is great that there are markers in place (the color circles and numbers), they are not unique and can as easily misguide the algorithm as to aid it. ie, the feature of "1" will be the same as the feature of "11" since to the algo, it just see 2 "1" instead of "11". Same for "6" and "9", "12", etc. Another issues, is the color circle are again repeating the same color already used by another marker, which serve to confuse the algo further.

In your case we recommend using unique and high contrast marker instead of the monotone color markers currently used. From our testing, using newspaper as background for the capture have being shown to improve reconstruction quality.