Images for Photogrammetry - Turntable method

photogrammetry
3d

(Summitchairs) #1

Hoping someone can help...

I'm hoping to capture similar sized objects to the image attached using the turntable method. Obviously this image does not have sufficient lighting which i'm currently in the processing of sourcing, nor does it have an effective 'featureless' background required for the Photogrammetry process.

I'm in the process of sourcing a green screen but as far as I can work out I need a huge screen. The object is 6ft away from the back wall (a specified minimum distance to stop green spill onto the model) and the camera (Pi cam V2) is at a sufficient distance away from the object to ensure I get the full object in frame.

However as you can see these distances leave a huge amount of background space needing to be filled with the green screen, I calculate a screen size of 16ft x 16ft . My questions is this, if I used a green screen the same size as the grey wall on the image (which would obviously create a green floor too!) could I just crop each and all of my images to get rid of the unwanted background to the sides of the image? I've read that cropping the images pre-processing is a big no no but doesn't explain why... As far as I can work out, as long as all images are cropped to the same size I don't see why this would effect processing? Might be an extremely obvious reason why so apologies if it is!

I would just test it, but I don't really want to buy a green screen if it turns out not to be big enough!

Thanks in advance!


(Alex Nan) #2

What program did you try ? (Autodesk Remake)
How many photos did you took ? (96)
How many angles ? (3 - low medium high - 32 photos each)
360 on all rounds ? (yes)
That would be my suggestion, without investing in a green screen. You can also mix up with black stripes on the background, I found out that it likes when it's not one color. I also believe that it automatically gives a better alignment for each photo.


(Mesheritage) #3

Well, you don't crop the object from the image because usually it uses the background to match the picture. However, if your background is not moving, this is not possible, hence you will have to crop anyway.

Adding a green screen might just do that automaticaly, but don't need to get bother with that, you can put any colour, the green is just a standard approach. Also you can just take a picture of the background and automatically erase it from the image. It is fairly simple if there is no conflict of colour (ie: background and object similar object).

Look like you have enought lifght, but not sure how the feet of the chair will work, it is very "plain black".

You could also add markers if needed, like on the black feet. Something you could correct in post processing.


(Summitchairs) #4

Hi,

Thanks for both your responses! Sorry for the delayed reply too!

AlexNan:
What program did I use? Reality Capture
I took 80 images at 22.5 degree increments across 360 degrees (i.e. at 16 different points). Now in terms of angles this is a bit of an unknown. I've produced a camera mount to try and automate the process but I think this might be hampering the compatibility of the photo's so i'm currently rethinking.

I have since bought a green screen and am currently in the process of producing a frame for it. There seems to be a lot of contradictory information out there with some people saying a featureless background is required and others saying this hinders alignment. I guess if needs be I could add these black stripes if I find the featureless green screen does not work.

Mesheritage:
As I've previously said, I'm using Reality Capture. The workflow for a turntable method seems to be undocumented when compared to Photoscan's, where the masking of images can be used during the alignment step, whereas masking can only be applied to the reconstruction step in Reality Capture. Furthermore Photoscan provides markers for the turntable method to physically tell the software common points on the object while its rotating. The main reason we went for Reality Capture was mostly cost related.

I'm hoping the green screen, like you said, will make the masking process far more automated, especially when it is lit evenly. I have just found this comment on an additional forum though which suggests masking during the alignment process may be possible? What software are you using?

To explain the masking in RealityCapture - even we do not have a tool for editing masks build in our app, the app itself supports masks to some extent. For alignment step all you need to do is to modify image color channel and paint it with a solid color. This will cause that natural features will not be detected there and thus will not influence camera alignment. For meshing we support alpha-channel masks. Adding alpha channel and masking only important parts actually speeds up whole computation. Another benefit is that you would not need to use reconstruction region to filter parts which are not important.

As regards to the feet of the chair, I agree this could be a potential problem. I may get some 3D scanning spray to matte the finish and try and light significantly without getting over exposed areas. Failing this I guess I could model externally and add to the mesh as a post process. I shall bare in mind also that I could add markers to the feet to help.


(Mesheritage) #5

Hi there,

It is interesting that you say "we choose RC for the cost". Usually we say that for cheap thing, not the expensive one! But I think it is better, they work differently and RC is just better if done properly.

Normally, you should not mask image for alignment, even for Photoscan (as noted in the manual). But this is when the background is changing with the object.
For any software you will have issue with fix background, because the software will try to match the feature in the background that might be stronger than the object. The "color" thing can tell the software that the "point in that area should not be considered".

You can align with mask, it will just rely on the feature of the object.

But to mask, it can be done automatically, independently from RC or photoscan. I am not sure if there are known software out there. I am playing with my own algo, but could be shared at one point.

To mark the turntable you can also cover it with feature surface, like "a newspaper". It would be better than a white surface, Then you can erase it postprocess. But it will be good to place the camera to some extend and maybe help out model the feet.

I am using research software (visualSFM, openMSV, bundler, openMVG, COLMAP). They are less interactive than commercial and sometime cannot be used for "commcerical purpose" but they are free and I am just doing research anyway.


(Summitchairs) #6

RC actually offers a 3 month license for €99. I'm hoping that once I've nailed the workflow, i'll be able to spend time photographing several products, then spend 3 months doing all the imaging processing.

I see, I'll have to test using masking during the alignment step and reconstruction steps and see the differences, however it sounds like I should avoid this for alignment. I'll also try and find options within RC to make sure it knows it is a static background and to use the colour to make sure it doesn't consider those points... is that what you meant? Sorry if I have misunderstood. Do you have any experience using RC with turntables? Any specific settings I should keep an eye out for?

As regards to the masking, I'm planning on using Photoshop. Although I haven't got much experience using it, my colleagues have, I believe you cant create an automatic process too using a Macro type feature.

The use of newspaper on the turntable is something I have heard before so maybe I shall go down that route!

I hope your research is going well!


(Mesheritage) #7

Well if you do not mask for the alignment, normally it will not work in the case where the background is not moving. I don't think you can avoid it.
There are algorithm to extract background, but I doubt they are implemanted in such software for now. Often they rely on background reference or colored like for the green screen. Not sure if RC can do that but it is not "ground breaking" so there is certainly some plugin out there.

I never used RC, I tried to help out regarding basic of photogrammetry but maybe there are some magic tools!
However, I see several people ending with a turn table "for the camera" rather than the object as they were not so happy with the optimum result. On the other hand I saw some compagny providing "turntable solution" with 0 assessment of their quality.

Also you are doing "large" objects, with different reflection properties. I think you might have to explore many option!

You can create automatic process, not sure on photoshop (I am not using it). What is doing done in image processing is quite advance.
But if you have a limited set of picture (not too large), maybe better to do it by hand as you can control the "performance" of the mask


(Summitchairs) #8

That's fine, was fully expecting to do masking at some point so hopefully the software plays ball during the alignment process. I have a couple of colleagues who are quite skilled at using Photoshop so hopefully the extraction of the background should be relatively simple, whether that be an automatic or manual process.

Struggling to figure out how the camera on the turntable would work, like some kind of rotating ring around the entire object? You able to provide the link to that company providing the turntable solution?

Yeah not too sure how effectively the object will scan but we shall have to see!

This is my first 3D scanning project so apologies for my lack of knowledge!


(Mesheritage) #9

The "turntable around the object" can be build DIY, there are various video on youtube about that.
I never encounter a compagny selling that stuff, they often turn the object and control the background. One went for the option of turning the camera but it was a "prototype" (and I forgot the name!).

Normally, if you have an option to colour the object of interest versus background, you should be able to consider the alpha of that.

You could as well take the picture around the object, without turn table. If you plan to add alpha from photoshop, the acquisition would be not much significant in term of time spending compare to PS. Might be a good training in the technique as well.

Good luck :wink:


(Summitchairs) #10

I shall have a hunt around on YouTube, would definitely like to see the 'turning of the camera around the object' prototype, that's originally the concept I was going for but decided against it!

I shall definitely be trying alternative methods too! Will be interesting to see the differences between the outputs using different techniques.

Thanks, really appreciate all your help!