Large scale RTIs? Stitching RTIs?

[BrilliantName]

I'm researching ways to stitch multiple RTIs of one artifact. I haven't found much info on the web.

I would like to have a very detailed scan of a larger artifact, more detail than I could get with the FOV that would fit the whole artifact in one photo.

I've seen some mention of stitching photos, which I'm very familiar with, but how would that fit into producing RTIs?

If that route is feasible, would one first stitch the raw photos and then generate the RTI? I see a lot of issues with that, as the light direction and angle would not be the same.

Maybe stitching final RTI would be a way to go, but I don't see any tools that would help?

Any suggestions?

[Dave Martin]

As you suspected, you definitely don't want to stitch images before processing.

One thread I've found is:
https://forums.culturalheritageimaging.org/topic/542-stitching-together-multiple-rtis/

I seem to recall a discussion (but can't find the thread right now) on merging RTIs of columns, but can't spot that thread right now.

As well as the long-standing RTI generator, there is the new 'Relight Lab' under development, would suggest you enquire there (sub-forum here as well).

Dave

 

[DeathMetalBeanbag]

I think there are several possibilities that might give you a usable product:

1. If the RTI data collection is very controlled and with a flat subject (as is mentioned by Carla Schroer in the post Dave linked) then it might work. I know that Klaus Wagensonner at Yale (see https://www.academia.edu/45614487/RTI_Stitching) has successfully done this and we had a brief email conversation about it.

Basically he suggested using a dome or dome skeleton setup to capture full overlapping RTI sequences for each section of the object (i.e. capture all light angles for one section without moving the camera then moving on to the next) and keeping the dome orientation and light positions exactly the same between image sequences. He then uses PTGui Pro software to create a template for aligning and merging all of the images (using tie points between the pictures) to combine those with the same lighting angle together. After he just repeats using that template for the other image groups to get a standard looking RTI dataset. The image stitching could probably be done with other software like Photoshop, it would just take longer. Lastly he processes the RTI with the light position data of the dome, not using the highlight method.

That's at least my understanding of his process; I still see some potential issues with it, particularly regarding the light distances changing across the scene, but it seems to have worked for him. I would encourage you to read through the linked presentation and maybe reach out to him!

 

2. I think the above could be repeated using the highlight method but you would have to have some way to get the lighting position in the exact same place for each image sequence which seems very difficult. I don't know the size of your object but maybe a much larger dome framework with enough room inside to change the camera position between sequences while keeping the light positions the same. But this might only be feasible in a macro photography setting... or with a gigantic light, polished bowling ball, and an old playground climbing dome!

 

3. Lastly, I personally have had good experiences just processing RTI normals images (highlight method) in segments with 20% overlap or more and then aligning them in photoshop. Taking that many RTI sequences of a large object can be painstaking (and it was) but the alignment could also be done manually if you wanted to avoid the extra image sequences to achieve the overlap. The trick was again keeping the lighting positions the same but for me this has been pretty forgiving considering I did it free-hand. I had the camera on rolling scaffolding, taking each image sequence with the same light positions, and moving the scaffolding for each new section.

I've attached an image of one of my subjects. It is a 4-5 ft. timber that I did in 5 sections. I will say it is not perfect as there are tiny inconsistencies in the normals that don't line up. However, my intention wasn't to use the normal data for anything other than visualization and I don't think it would work to create an interactive RTI in the viewer but then again I haven't tried. My light position accuracy between segments was good enough to get a good normals image but was likely not precise enough to merge the images together before processing like Klaus.

 

Hopefully this at least gives you ideas!

[DeathMetalBeanbag]

Here is the image of my stitched subject (quality very low due to upload limits but it gives you an idea of what my post-processing stitching looks like). There are really no seams present and the normals are quite accurate compared to the actual geometry of the timber surface all things considered. If anyone copies/references this please credit: Ian Dunshee, East Carolina University Program in Maritime Studies. Thanks!

[Dave Martin]

Ian, that's brilliant, thanks for sharing.