Agisoft Metashape
Agisoft Metashape => Face and Body Scanning => Topic started by: Magnus on January 21, 2013, 02:55:18 AM
-
Hello all!
I had not initially planned on posting the testing I've done until I'd gotten more proper results but I thought that it might be interesting for you to see the crude progression, hehe.
The goal I had for this was to be able to capture a head and shoulder portrait in 3D, at a quality of acceptable level. Later I plan to progress to a fullbody capture solution.
IR, Lee, has already provided a lot of the information needed thru his experimentiation, but I thought it would be a good learning experience to try some different ideas I had.
The setup that I started with consisted of 2 600D and 4 1100D. I used Smart Shooter for the transfer of the files to computer and for checking focus etc. and a simple corded remote trigger connected thru splitters to the cameras. I used "11 inch articulating arms" to be able to attach and easily adjust the camerapositions. For lighting I used this LED light http://www.leds.de/en/LED-lamps-and-luminaires/Ultraslim-LED-Panels/Eco-LED-Panel-coldwhite.html (http://www.leds.de/en/LED-lamps-and-luminaires/Ultraslim-LED-Panels/Eco-LED-Panel-coldwhite.html)
For the photos in the examples posted I used F14, 1/25th shutter and ISO400 on all the cameras.
An important thing to note was that I shot in JPEG (since I was mostly just testing the multicamera mode on Smart Shooter and checking coverage) not in RAW so a lot more information is possible to be retrieved. The compression in the 1100D is quite harsh and removes a lot of the detail information. So keep this in mind.
So in this first post I attach a picture of the rig (picture taken with mobile phone so please excuse the poor quality).
More to follow.
Best, Magnus.
-
With this second post I have attached an image where you can see the distances of the cameras and of them to the subject.
As you can see they do not in this test provide that much circular coverage.
I currently have 2 600D and 8 1100D (plus some other cams I can throw in) they are not yet set up in a rig though.
-
In this third post I have attached the results from that setup.
I processed at "Smooth" and at "Ultra High" with 1M faces.
You will notice that it is noisy but with a decent amount of detail (again keep in mind this is processed from JPEG and not RAW).
-
Here's another image at the same settings except this is at "High". In this you can also see the processing time.
-
Will be watching every step Magnus...
-
Very interesting tests Magnus, great to see you sharing setups, ideas and results.
From my own tests with RAW and Jpg I can honestly say there is zero difference in quality. The only 'bonus' you get is a 5-10x increase in loading and processing times! because all images are much larger in file size. It added zero extra detail in my tests. Plus colour textures are slightly more detailed, less compression.
You could probably improve quality by not using ISO 400, this is quite high. Ideally you want ISO 100 or below! as high ISO introduces fake noise data that wont exist in other photos.
-
Not meaning to hijack at all, but just want to show what happens with 48x DSLR and flash light.
Exp 1/4, AV 12, ISO 100
HIGH at 7.5 million. 360 degree.
(http://www.ir-ltd.net/uploads/42-cam-head.jpg)
The bonus of Flash is very bright light at super fast speeds. As you are effectively synchronizing at around 1/10,000th speed. Which means you can use lower ISO. Unless you use super bright constant illumination. Depends on how fast and bright! your lights are.
-
I have to agree with Lee also..... Low ISO, bright lighting and sharp images are the key to success.
Greetings,
Mr. Curious
-
I agree with Mr. Curious but want to stress that "bright lighting" is too simple: ideally, the lighting should be such that small detail (e.g. skin pores in this case) will be enhanced rather than subdued in the images. Visually smooth surfaces are really difficult because PhotoScan needs to be able to detect features. Experimentation with multi-directional vs. homogenous diffuse lighting might be interesting in this respect.
-
I agree with Mr. Curious but want to stress that "bright lighting" is too simple: ideally, the lighting should be such that small detail (e.g. skin pores in this case) will be enhanced rather than subdued in the images. Visually smooth surfaces are really difficult because PhotoScan needs to be able to detect features. Experimentation with multi-directional vs. homogenous diffuse lighting might be interesting in this respect.
The reason bright light is highlighted here is because of the experimentation with continuous light. To mimic the same light levels one gets with flash light which is VERY bright but at a short burst of time, anything around 1/10,000th of a second. To match that same level of light to use similar camera settings, low ISO (very important) but high exposure speed 1/100th you need bright light just to match that same quality. Obviously not over bright as this will wash out any details. The alternative is noise projection but this introduces other problems of being able to capture a color pass quickly straight after, or using additional separate texture cameras. Multi directional lighting is possible but you need VERY fast capture if you are doing scanning of live subjects.
-
I agree with Mr. Curious but want to stress that "bright lighting" is too simple: ideally, the lighting should be such that small detail (e.g. skin pores in this case) will be enhanced rather than subdued in the images. Visually smooth surfaces are really difficult because PhotoScan needs to be able to detect features. Experimentation with multi-directional vs. homogenous diffuse lighting might be interesting in this respect.
The reason bright light is highlighted here is because of the experimentation with continuous light. To mimic the same light levels one gets with flash light which is VERY bright but at a short burst of time, anything around 1/10,000th of a second. To match that same level of light to use similar camera settings, low ISO (very important) but high exposure speed 1/100th you need bright light just to match that same quality. Obviously not over bright as this will wash out any details. The alternative is noise projection but this introduces other problems of being able to capture a color pass quickly straight after, or using additional separate texture cameras. Multi directional lighting is possible but you need VERY fast capture if you are doing scanning of live subjects.
Hello,
I agree my response of saying just "bright lighting" was an over simplification on my part. However, Lee did a perfect job of translating what I intended to say between the lines. Thanks Lee, you put things far better than I could have!
Greetings,
Mr. Curious
-
Yes, I agree. I was still mainly concerned with the initial problem of model noise appearing on a smooth face and not so much with the problems of photographing a 3D scene of moving subjects.
-
Hello!
Thanks for all the input guys!
Lee, the reason I assumed RAW to be a step up for the 1100D in particular is that in the pics I took and compared, the Jpgs had lost much of the skin detail and were a lot softer. Also since I am on the lower end with regards to MP, my thought was getting as much detail as possible.
It might also be the sharpening settings on the camera as I've read that they are set pretty low.
Have you tried RAW vs Jpg at Ultra High setting?
Ofcourse as you say with those kind of added "bonuses" it might not be worth looking at, hehe.
I am very thankful for all your suggestions so please "hijack" all you want. ;)
Your results just helps to show the power of a complete rig and that is very valuable.
I think my experiments here will mostly show the pitfalls there are, but that's exactly why I do it, so others might avoid them. :)
I did actually use the Canon SpeedLite 580EX II on that same rig, bounced away into the wall and ceiling and it was ofcourse illuminating the subject better (more evenly) than the LED, hehe.
A point about noise projection, I was actually testing that with a mannequin, it ofcourse having no skin detail. What I did was changing the projected image from a noise pattern to a white background so I got both a noise and color image. Then I just exchanged them after having constructed the model and generated the texture from the latter images. Quite a fun experiment.
Also another reason why I am, in the beginning, considering a rig with fewer cameras and adding more later is the excellent example here http://www.ten24.info/index.php/3d-scanning-on-a-budget/ (http://www.ten24.info/index.php/3d-scanning-on-a-budget/). My thought was less investment on cameras and lesser computing to construct a model and a bit more time doing manual labour on it in the beginning whilst starting up. Then add more cameras when business grows.
Best, Magnus.
-
Hello!
Thanks for all the input guys!
Lee, the reason I assumed RAW to be a step up for the 1100D in particular is that in the pics I took and compared, the Jpgs had lost much of the skin detail and were a lot softer. Also since I am on the lower end with regards to MP, my thought was getting as much detail as possible.
It might also be the sharpening settings on the camera as I've read that they are set pretty low.
Have you tried RAW vs Jpg at Ultra High setting?
Ofcourse as you say with those kind of added "bonuses" it might not be worth looking at, hehe.
I am very thankful for all your suggestions so please "hijack" all you want. ;)
Your results just helps to show the power of a complete rig and that is very valuable.
I think my experiments here will mostly show the pitfalls there are, but that's exactly why I do it, so others might avoid them. :)
I did actually use the Canon SpeedLite 580EX II on that same rig, bounced away into the wall and ceiling and it was ofcourse illuminating the subject better (more evenly) than the LED, hehe.
A point about noise projection, I was actually testing that with a mannequin, it ofcourse having no skin detail. What I did was changing the projected image from a noise pattern to a white background so I got both a noise and color image. Then I just exchanged them after having constructed the model and generated the texture from the latter images. Quite a fun experiment.
Also another reason why I am, in the beginning, considering a rig with fewer cameras and adding more later is the excellent example here http://www.ten24.info/index.php/3d-scanning-on-a-budget/ (http://www.ten24.info/index.php/3d-scanning-on-a-budget/). My thought was less investment on cameras and lesser computing to construct a model and a bit more time doing manual labour on it in the beginning whilst starting up. Then add more cameras when business grows.
Best, Magnus.
Yes that's a good method, start small and build up.
I haven't tried processing RAW at Ultra no. I think if/once we get Network support running allot of tests in Ultra will be easy to then determine. 5-10 hour waiting times is no good at the moment. Plus I think 64GB in that instance is not enough on one machine either at 70-90 images at 18MP. So lets see if these talented guys at Agisoft implement networking.
RAW does seem to retain more detail information (as long as you are in optimal focus range!) but it doesn't do much at High. BUT. If I get 3-4 hour processing in RAW than compared to 1 hour at HIGH. I would imagine RAW at ULTRA would yield 24 hours or more!
-
Hello!
Network support sounds like it would open up a lot of opportunities.
I have a question regarding cameras for you. Would you recommend going with fewer but higher MP cameras, like the D800, or is it preferable for the geometry reconstruction to go with more lower MP cameras so as to get more angles covered?
This is only very slightly related to this topic but I thought it so interesting that I wanted to share it anyway. It is this http://www.untitledfilms.com.au/blog/2012/12/micro-expression-exploring-motion-image-photography/ (http://www.untitledfilms.com.au/blog/2012/12/micro-expression-exploring-motion-image-photography/) about the Canon 1D C that shoots 4k video. Those stills are quite impressive.
Best, Magnus.
-
Hello!
Network support sounds like it would open up a lot of opportunities.
I have a question regarding cameras for you. Would you recommend going with fewer but higher MP cameras, like the D800, or is it preferable for the geometry reconstruction to go with more lower MP cameras so as to get more angles covered?
This is only very slightly related to this topic but I thought it so interesting that I wanted to share it anyway. It is this http://www.untitledfilms.com.au/blog/2012/12/micro-expression-exploring-motion-image-photography/ (http://www.untitledfilms.com.au/blog/2012/12/micro-expression-exploring-motion-image-photography/) about the Canon 1D C that shoots 4k video. Those stills are quite impressive.
Best, Magnus.
I can't say but I personally don't like working with full frame cameras as I am used to framing and using DX in the Canon 550 and 600D. You have to start using more restrictive lenses in FX mode. I have 6x D800's but not enough to really test to compare against an array of 18MP Canons.
The 1D looks good but it's very expensive also rolling shutter I believe like all of the range? you could build a much better dedicated MV camera array (global shutter) for less I think and with more custom support.
-
Hello!
Ah yes, FF is great for creative shooting but not so much for what we are doing here. Your plan is mostly to use the D800 for high res texture extraction?
I was checking out open source camera project Apertus http://www.apertus.org/ (http://www.apertus.org/) The image sensor they are planning to use is quite interesting http://axiom.apertus.org/index.php?site=imagesensor (http://axiom.apertus.org/index.php?site=imagesensor) Cool project to keep an eye on!
Best, Magnus.
-
Hi Magnus and Infinite,
Are the issues you describe with the full frame sensor just specific to this application i.e. close range/body scanning?
I had always been under the impression that the larger sensor would be advantageous for photogrammetric applications, in my case 'medium range' - mostly buildings and other large structures.
I had for a while been pushing to upgrade our D7000 (DX) to a D800 (FX) on the assumption that it would be an improvement!
James
-
Hello James,
full format sensors usually have better image quality (e.g. less sensor noise and less "bleeding" between pixels). For close-range applications they are often not advantagous because at the same field of view (larger sensor in combination with a longer focal length) they have a smaller depth of focus. For facades of buildings, air photos etc., full format sensors would be preferable, for close-range applications you'd have to find a compromise between image quality and depth of focus.