I’m evaluating several photogrammetry software for capturing retail store interiors and based on my testing couple limitations I’m seeing would prevent me from using this for production. And would like to know if there are any workarounds to this.
- It seems like photogrammetry in general are designed for outside-in capture such as sculptures, dolls, etc. Rather than inside out capture of environment such as store and rooms. Is this correct?
- All the softwares I’ve tested had problems with reflective and transparent surfaces. Is there any way around this?
Rooms are no problem if you take the pictures properly.
You want lots of diffuse light in the room for once.
Be aware that flat, featureless surfaces (like a white wall) will cause problems. So add something to it. Markers, post.it notes, project a pattern, …
As for shiny, reflective and transparent objects or surfaces:
Those are an inherent weakness of photogrammetry and all photogrammetry solutions have their troubles with them. Incorporating laser scans in your workflow may help with those.
When taking pictures of a room I would advise to move around the room, back to the walls, photographing the opposite side of the room. Do it multiple times from different heights and angles.
Even in the best of cases you will have to invest quite a bit of manual labour in post processing though.
Some recommendations to taking pictures: Taking pictures for photogrammetry
I can confirm, interior is not at all a problem.
A laserscanner has the same weakness though come to reflective surfaces, so that won’t really help in this case.
And retail is bound to have many unicolor and reflective surfaces…
Spraypainting them with easily removeable paint is one option, but I am not sure if that is manageable here… 
Thanks for the replies. But I have not found single example of good interior photogrammetry other than simple flat wall warehouse type. I found a few apartments and bathrooms which didn’t have the quality I’m looking for. Can you share some good examples of indoor photogrammetry?
And I did take the pictures based on the guidelines and back to the walls taking pictures across the store in multiple loops. In various heights and angles. Still can’t get very good result. It captures the objects within the store pretty good but the walls, floor and ceiling looks to require a lot of clean up.
Obviously, what we’re evaluating this technique against is with traditional modeling. If the clean up process is more time consuming than traditional modeling, this obviously wouldn’t be worth it.
lubenko posted a link to some indoor project over in the FAQ section of the forum.
I’ll repost it here as it bears some significance regarding your questions.
https://developer.valvesoftware.com/wik … ogrammetry
Ultimately you will end up with a mix of photogrammetry and modeling.
ShadowTail wrote:
lubenko posted a link to some indoor project over in the FAQ section of the forum.
I’ll repost it here as it bears some significance regarding your questions.
https://developer.valvesoftware.com/wik … ogrammetry
Ultimately you will end up with a mix of photogrammetry and modeling.
Thanks, but I’ve seen that one and that just confirms my initial assessment. For such simple interior, initial capture doesn’t seem very clean compared to object capture examples. And seems like a lot of clean up work. I’m wasn’t saying photogrammetry is not possible for interiors. It just doesn’t seem as effective as capturing objects.
Brian, I think your statement is a bit too general.
You probably have more or less round, simple objects with a distinct texture in mind (rocks).
Comparing those to an intricate interior (the link doesn’t work atm) is apples and oranges.
If the interior is simple, it is not gibt deal.
If the object is very sophisticated with reflecting materials, it is very difficult.
I’m sure you’ve seen the golden carriage here in the showcase.
Same is true for interior.
I attached a project that I am currently working on.
interior_1.jpg
Texture is ideal:
interior_2.jpg
However, geometry is also not too simple - I removet the altar part because it is not relevant in my case.
interior_3.jpg
It is noteworthy that I did NOT take the images for the vault itself, but only for reference of capitals and bases, which are not visible here.
So in your case I think it is simply a question of effort compared to result.
If what you describe as retail is very modern shops with loads of glas shelves, then modeling will be much quicker, especially since the geometry is rather straightforward.
Thanks for sharing that. It looks great. What we’re evaluating is all the stores in a shopping mall. So it’s all kinds of stores, fashion, jewelry, restaurant, electronics… The glass shelves, we can model, but it’s the merchandising inside the shelves that will be the challenge. Also the walls that have merchandising on didn’t capture very well.
We’re just looking for best possible solution.
For now, we’re looking into modeling the walls, floors, and ceiling using traditional modeling and the merchandising using photogrammetry as it seems more efficient at capturing free standing objects.
Transparency… still not sure.
This is an equirectangular render (360x180 deg) from a photogrammetry model I processed for a friend… sorry, can’t show you anything bigger than this.
The obvious is standing right in front of you… you don’t just have a room, you have a room and several objects so you aren’t going to get a good model of everything using a strategy for an “empty” room.
14mm rectilinear lens, multiple camera heights … at least 2, probably 3 for a shop where they stack things higher than “normal” rooms… 50cm, 1.2m, 2.2m high… and a bucket load of shots. Depending on the free space between displays you’re looking at a camera spacing of at most 50cm, probably less. AN additional set of images shot along the free walking areas looking down at the floor, and a grid of shots to capture the ceiling/upper sections of wall. If you want better details in objects displayed on tables then you’ll need to treat each table as a large object and shoot accordingly.
I try to teach people to think about their shoot purely in terms of the surfaces they want to capture and then shoot for that.
Very nice manual, Ben!
Most people (including myself) just never get their heads around the sheer amount of images one has to take.
And that this is the most time-saving factor of all!
I think everybody has to go back 3 times and spend 4 times the hours in trying to fix a model before appreciating that! :lol:
Again, myself included.
What never really works for me is the individual surface approach, because then there are often misalignments due to steep angles from other images sets, that are intended to cover other areas. How does one avoid that? Maybe it is just due to my (now remedied) reluctance to use RAW data.
Basically what happens is that I take general shots (I manage that ok) but if I have to zoom in to cover some detail or obscured part, I end up with steps due to misalignment: Manual improvement of alignment / sparce point cloud
Any idea what I could do better?
Ah yes… forgot my other tip… never stop taking photos. the other common problem is to stop shooting, move somewhere else and start again with the risk that there isn’t enough overlap somewhere or the jump in magnification is too great. The extra shots to ensure connections can greatly reduce the amount of post work to join multiple components. The software may be able to link photos from anywhere in the image set but it certainly doesn’t hurt to guide it along the way.
Hehe.
The moving thing is what I did after my first try, which was the typical stand-in-one-position-and-shoot-circular mistake.
Although considering that I did get acceptable results. Only the details were problematic.
I like to imagine taking a video covering any visible surface and tying all of the paths together like a cobweb all over the scene.
But still I end up with alignment issues.
Anyway, RAW will show with the next project if this was the issue…
