Room Capture Only using 4 images of 300 In Image alignment.

Answered

Comments

44 comments

  • Avatar
    Erik Kubiňan CR

    Hello dear user,

    Could you please share with us some screenshots of the interface to let us see images components and the 3D viewport. Also, please post here a few images you've captured so we can see what exactly are you shooting.

    1
    Comment actions Permalink
  • Avatar
    daniel.osullivan

    I have tried the image overlap setting on low medium and high and every time I click align images I get a different result. 

     

    https://i.imgur.com/SOf3pFY.png

    https://i.imgur.com/Y9VVjB6.png

    https://i.imgur.com/xZlFBaC.png

     

    I went in a circle around the room shooting inwards. I then also did a loop of the table in the center of the room.

    0
    Comment actions Permalink
  • Avatar
    Erik Kubiňan CR

    The settings are okay, but the room has white walls and the carpet is almost featureless. RealityCapture needs to see detail/structure/pattern in your images to align them together. A white wall doesn't provide RC with any points to detect as well as the carpet. Although, you should be able to create at least something out of it and the wooden stuff and the chairs should be reconstructed aswell. Please try adding more images and move around the room while shooting to get different perspectives of the subjects.

    0
    Comment actions Permalink
  • Avatar
    Jonathan_Tanant

    Are you sure you do not have a larger component? can you share the component list? (left pane in "1D")

     

    0
    Comment actions Permalink
  • Avatar
    daniel.osullivan

    Yeah here is it. https://imgur.com/a/fiW2vXA

    0
    Comment actions Permalink
  • Avatar
    Jonathan_Tanant

    Ok, so you seem to have a 156 pictures component aligned here. You even have 2 models reconstructed in this component. This does not look so bad, much better than the 4 at most that you say you are having, unless I misunderstood something?

    0
    Comment actions Permalink
  • Avatar
    daniel.osullivan

    Yeah thats the most ive managed to get alligned but it literally just came from hitting allign imageses like 5 or 6 times it gives me different results each time. Still not enough to make a decent mesh from.

    0
    Comment actions Permalink
  • Avatar
    Mike Simone

    Alignment can be somewhat of an art, and highly depends on your source photos if you want a VERY precise alignment.

    There are guides on how to improve it on here somewhere if you search.

    0
    Comment actions Permalink
  • Avatar
    Erik Kubiňan CR

    daniel.osullivan Alignment improves after each try it is not a random occasion the accuracy increases after realigning and as Mike Simone said you will need to improve your shooting if you want to get a perfect alignment, it is just a practise, nothing to be scared of.

    0
    Comment actions Permalink
  • Avatar
    daniel.osullivan

    I’ve taken into account trying to get multiple feature points in the photography. I’ve given it 500 new images this time and I can usually only get 2 match points Agisoft gives me a very decent model.  Where as this give me 2 matches. I must be doing something very wrong to be getting such awful results, but I have redownloaded the software multiple times.

    I can’t understand this as I’ve seen people get amazing results with this application. Anything I do comes out like trash.

     

    blob:https://imgur.com/d7b96ca7-c8ab-42d9-8c63-3a12414fcb6b

     

    Like do I have to select all the images before I align there must be something obvious im missing. I just load in the photos and hit align.

    0
    Comment actions Permalink
  • Avatar
    daniel.osullivan

    Erik Kubiňan CR That doesnt seem to be the case for me. One of my runs it managed to get about 200 of the photos and add tie points. Then I hit align photos again and i was back to only 5 photos with tie points.

     

    Is there anyway to tell it which photots to start from when alligning images or does it chose images by random?

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    They make it sound easy - just improve your photoing technique! I found after months of trying everything with building interiors, really no luck. Even with quite knobbly cracked old plaster RC couldn't find anything to consistently catch onto. Not only 'featureless' areas, but linear objects too, where one bit of the linear was hard to distinguish from the next.

    To me, this spoiled the building-interiors use of RC that I was hoping for, so gave up. Tho other folks like Götz Echtenacher seem to make a professional success of it, it defeated me.

    The only way I could get building interiors photos to align reliably was to pin screwed-up balls of colourful patterned wrapping paper to the surfaces at say 50cm spacing. Then RC could see and distinguish several in each photo and could align to them - and the plain surfaces between were also happy. But that was much too slow and laborious for professional use, as a fast alternative to traditional measured survey of building interiors.

    It's worth mentioning that Alignment is only for the purpose of aligning the photos - unique to RC AFAIK, the resultant sparse point cloud is not used at all for subsequent stages - once the photos are aligned RC starts all over again to construct the mesh (and later texture it) by other means.

    0
    Comment actions Permalink
  • Avatar
    daniel.osullivan

    Tom Foster Yeah i've actually printed out some a4 pages with different shapes that i'm going to stick on the walls and floor. It's just very weird that sometimes the image alignment works so well then other times I only am getting like 4 or so. I basically just have to sit and hit the align image button until I get a decent result. 

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    Trouble is, that screws up the images for the texturing stage.

    It's not just hitting the Align button repeatedly - there are other things to try e.g. we're told to delete all Components but the best one before re-trying Align. And there are non-default recommended settings for Align. And more - lots of advice, going back years, to search for about problematic Alignment, on here.

    If you or anyone can come up with a straightforward workflow for building interiors, I would be agog to hear about it - please!!!

    0
    Comment actions Permalink
  • Avatar
    Erik Kubiňan CR

    Dear users, 

    Speaking of clean interiors the main successful method is using either laser scans or a combination of photogrammetry and laser scans. As some of you cannot afford a laser scan you can always try using just photogrammetry but you need to understand that shooting a clean surface mostly featureless won't always end up well, so this is not an issue of the alignment process itself as a subject should be picked well first. Anyways as you guys have noted yes you can use some patterned papers, stickers,etc. to make for some features and get a successful alignment. When using this technique you can as well shoot the interior clean and then shoot it again without moving any objects with the "stickers". Then once you're happy with the result you can use the clean images to texture it. Image layers function can be put to a good use here for example, or you can do it manually, that's up to you.

    0
    Comment actions Permalink
  • Avatar
    Mike Simone

    The absolute best way is to just use a program like Maya to clean up the model and turn millions of polygons to a few hundred.

    That's the absolute best and prefered way imo.

    But there is a ton of post work

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    Erik, that is extremely helpful, for a RC representative (?) to say clearly that laser scan is the straightforward way with 'clean' interiors. In years, I haven't seen anyone say that clearly on this forum. So now we know!

    It's also disappointing, as photogrammetry held the promise of ultra-fast capture with simple portable equipment (camera), as well as the ability to capture every hidden corner, which laser scanning from few station-positions can't.

    It's also puzzling, as a succesful industry captures interiors for e.g. estate agents at high speed and low cost, by the SLAM version of photogrammetry, Paracosm being the latest arrival.

    I still think sometimes I should offer to work as slave-assistant-apprentice for e.g. Götz Echtenacher, to learn his secrets with interiors!

    Mike, post-processing as you describle is absolutely prohibitive, from my perspective.

    The new building industry 'BIM Alliance' led by Bricsys CAD within Hexagon Corp which also owns Leica, plus collaborating architecture and engineering cos, aims to make 'Scan to BIM' a practical reality, whereby exterior and interior scans combine to leave a void volume between, which is the structure of the building - and AI (which Bricsys pioneers) then creates as CAD solids and designates as walls, floors, ceiling, roofs etc - all automatically. So far, only Leica's speciality, laser scanning, is envisaged as scan source, while other CAD/BIM cos see photogrammetry as equally important i.e. mixed sources combined, as Erik suggests

    0
    Comment actions Permalink
  • Avatar
    daniel.osullivan

    Just noticed that everything is getting split up into components. Each component has about 3 or 4 cameras in it. I doubt merging them together would do anything as they are orientated and posiitiond all differently from each other.

    0
    Comment actions Permalink
  • Avatar
    Jonathan_Tanant

    How you work with components is on of the most import things in RC. You should read the doc to perfectly understand this.

    Then the method is to align, delete the small components (<20 pictures), tweak the settings, possibly add control points and align again... until you have a good large components.

     

    0
    Comment actions Permalink
  • Avatar
    Jonathan_Tanant

    I don't want to start a debate about laser scanning VS photogrammetry, but... (those are only a few thoughts)

    I can't totally agree with the "laser scans are better indoor". It really depends on the features you have (how textured your surfaces are) - and photogrammetry keeps very important benefits, like :

    -cheapness (a 3k€/$ camera and lens VS a 40k€/$ laser scanner - even with a 500€/$ camera or smartphone you can have very decent results).

    -weight (a laser scanner is more heavy to operate - a camera can be as light as less than 1kg while the laser scanner will be something like 5kg without the tripod. and can't be operated handheld).

    -with a laser scanner you can't operate discreetly, while with a camera you are just seen as a tourist taking pictures (a lot of pictures ;-) ).

    -there is something interesting with the fact that with photogrammetry an improvement of the software part can improve the result, with the same set of pictures. With a laser scanner you have the quality you have, and that is it. 10 years after this is still the same point cloud.

    -just speaking about the quality, this is true that on feature-less surfaces the scanner will do a much better job. But if you have more features, the photogrammetry can be better I think. It really depends on the features you have in your scene.

     

     

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    It's strange that even ultra-high-res photography (assuming sharpness already) doesn't seem to significantly improve RC's ability to detect small features in flat plaster - if only paintbrush marks - which are always there, clearly visible from say 600mm away.

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    Erik Kubiňan CR  said:

    "you can as well shoot the interior clean and then shoot it again without moving any objects with the "stickers". Then once you're happy with the result you can use the clean images to texture it."

    I don't get that. Don't the 'without stickers' set of photos have to be taken from the identical position, without moving the camera, as each shot of the 'with stickers' set? That wd be completely impractical.

    If the stickers enable the first set of photos to Align, doesn't the second set also have to be Aligned also, for RC to use them just for texture imagery? If the second set is taken by going round a second time with different camera positions, how can RC decide what bit of imagery to use where, if the second set is not Aligned?

    0
    Comment actions Permalink
  • Avatar
    Jonathan_Tanant

    Everything has to be aligned (with and without stickers), then you disable the stickers pictures for texturing.

    The stickers will help to "connect" images together. They all need to be unique and full of features, to avoid false connections.

    I used chromatic noise pattern video projection and it really helps to get precise surfaces reconstructions.

     

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    "chromatic noise pattern video projection" - whattt? Sounds good - can you give a hint in 'for dummies' language?

    "Everything has to be aligned (with and without stickers) .... The stickers will help to "connect" images together." Do you mean that once the 'with stickers' pics are Aligned, RC has a better chance of recognising weak features in the 'without stickers' pics and can Align them with the 'with stickers' ones, even tho taken from different camera positions? So no one-to-one correspondence between the two sets of photos is necessary? Would in fact fewer 'without stickers' pics be OK, covering everything but with less fanatical overlap?

    0
    Comment actions Permalink
  • Avatar
    Erik Kubiňan CR

    The stickers/non stickers images don't have to share the same camera positions.There should be some key features in the interior (table chair etc) literally anything every now and then around the interior which will be recognised as the same feature as on stickers photos so you will be able to align both photo sets together.

    0
    Comment actions Permalink
  • Avatar
    Jonathan_Tanant

    The chromatic noise pattern is a simple pattern of squares with random colors, which I project in the space using a short throw video projector (0.5 throw distance lens). And this helps a lot in the reconstruction of surfaces. 

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    Erik, that is fantastic clarification - thanks.

    Would my question "Would in fact fewer 'without stickers' pics be OK, covering everything but with less fanatical overlap?" be true?

    Jonathan, that looks brilliant too.

    "projector with 0.5 throw distance lens" - why necessary, and how does that work? Is that really close-up to the wall? Sounds like the projector is intended to be moved around as photography progresses? Is the projector close-up so as to project a really bright image? How would it work in a daylit room - pattern still visible? How would it work with flash as the main exposure-light source?

    0
    Comment actions Permalink
  • Avatar
    daniel.osullivan

    https://i.imgur.com/QASfo17.png

    Ive been getting some good results now using control points and deleting components that only contain 5 or so images. Ive managed to get a pretty good point cloud but its still missing all the shots from one of the 4 walls. I was using these image markers on an A4 page stuck around the place. If im going to do another reshoot i think i'd print out the image markers bigger and also use images that are unique from all angles like the suit spades as opposed to a circle or square. Also maybe adding color would help.

    https://imgur.com/D3Eay6w.png

    0
    Comment actions Permalink
  • Avatar
    Tom Foster

    That's why I used screwed-up balls, say 40mm across, of colourful patterned wrapping paper pinned to the surfaces at say 50cm spacing. With fine nail/pin/brads and a manual pin-push tool such as

    https://www.amazon.co.uk/Expo-75110-2mm-PIN-PUSHER/dp/B0053WW4GS

    I could place 3D manual markers v fast, unique in both complex shape and coloured pattern, onto both wood and plaster, damage almost invisible, and pull them off again by hand, leaving the pin stuck in the paper, ready for next use. No need for huge A4 targets, sticky tape.

    0
    Comment actions Permalink
  • Avatar
    Erik Kubiňan CR

    Tom Foster Well possibly but I wouldn't go too far with lowering the amount there still should be a decent amount of pictures since you're going to both align these and texture from them. You always have to try it first as I wrote there have to be at least some good features making the match of both photo sets recognisable.

    0
    Comment actions Permalink

Please sign in to leave a comment.