Share Gear / Software or Hardware Comparisons / Scan Experiments


(Banalesmuseovirtual) #16

Here is the translation (shortened):

Photogrammetric scans of Heritage elements are not always placed at an accessible high level. Photos of the upper sides are annoying and complicated, and it’s frustrating when the model is unfinished. You can find solutions to raise the camera position, but commercial poles for photography are expensive and I did not find any over 3.5 m. Drones, cranes and mobile platforms are even more expensive, and you need licence, permits, planning…

So this is can be an easy and cheap solution.

All the components can be found at a reasonable price, and you only need a drill for the upper hole. The total budget is 45 € (plus 45 € for the WiFi connection, that has other utilities.

But most important: does it work? It provides a maximum height of 4.5 m. I have used it several times, getting good results for the faces that are facing up. So this “top advanced mega-future sofisticated technology device” can be useful for anyone in this situation.

Sorry for my English! :wink:


(Banalesmuseovirtual) #17

By the way: another Sketchfab mate from Spain told me yesterday about a solution with a fishing rod and action camera that allows him to scan up to 10 m. height. I might test it this summer...


(Banalesmuseovirtual) #18

I just remembered this montage I did some time ago about my capturing kit:


I have added some more gadgets since then, but this can be a good start point that covers all kind of pieces to scan.
Hope it is useful!

Edit: I just realised of an error: the camera is obviously DSLR, not SLR.


(Misterdevious) #19

Just a noob photography tip for those that may not know:

Good depth of field (DoF) is imperative for scanning small objects. Good DoF will mean less artifacts in your final mesh, much less to clean up in post and the highest quality texture/image maps. Using a 50mm lens on my Canon 6D, I usually set the f-stop to f11 or greater to achieve 6" - 12" range of focus on the object (see the Simon scan for an example). This will usually mean that your exposure time will be 1/4 of a second or longer to get the best image possible. There was almost no cleanup of the original model (however, with this one I played with retopology) and the textures were sharp and crisp over the entire mesh. (I can post the original scan mesh for comparison if you like)

Anyway, that's my little tip... :smile:

Simon (Doll Head) by misterdevious on Sketchfab


(Nebulousflynn) #20

good advice @misterdevious - specifics like this are really useful. How big was the doll's head IRL?

would be great to see the original scan to compare if you get time to upload...


(Misterdevious) #21

Simon Scan (comparison) by misterdevious on Sketchfab

@nebulousflynn

This is a slightly decimated scan of Simon. 500k faces and a single image map at 4096x4096. I couldn’t upload the full 1,211 Million faced, raw scan. It got errors during the upload. This was decimated to 500k faces in Photoscan (to help the upload process and keep the scan geometry), then re-oriented in Blender. No other changes to the mesh had been made.

The scan consisted of 81 photographs using a Canon 6D, 50mm lens at 10 sec exposure (f22) - It’s a bit overkill, I admit… but good results.

The head measures 4.25” (10.8cm) in diameter (ear tip to ear tip) and 5.5” (14cm) tall and 4.75” (12cm)deep.

*edited for redundant info


(Nebulousflynn) #22

awesome work - thanks for sharing!


(Misterdevious) #23

Cheers!


(Banalesmuseovirtual) #24

Just remembered this comparison I did with the same photo set I took at the Museo Arqueológico de Córdoba (about 25 pictures, "museum-visitor limited" :disappointed_relieved: ) processed with Agisoft Photoscan (AP) and Reality Capture (RC). They were both retouched, remeshed and texture-and-normals baked in Blender to reduce to a simmilar number of poligons.

Museo Arqueológico de Córdoba by Pablo Serrano Basterra on Sketchfab

The resulting quality seems quite simmilar to me. I see some advantages and handicaps on both software packs:
- Precission: RC gives a bit more accurate geometry, specially when the photos are not optimal (better gestion of shades, shine, gaps, etc.)
- Speed: RC is much faster. I have obtained a complete model of more that 30 million faces from 700 fine quality photos in about 2 hours (try that with AP!)
- Price: RC is quite more expensive, though you can buy a basic (but quite complete) version for 3 months at only 99 $, so it is quite flexible.
- Interface and workflow: RC has been designed by Czech people, but they seem to be Martians! :smiley: It took me 1 month to understand their windows a little bit. Then, you can change every single value for calculations, but...
- Help: ... but there is not a good help function, any user's manual or simmilar. They have a forum to ask any question, and they are quite fast to response, but it is a bit annoying to need to ask everytime.
- Chunks and point marking: AP is a bit more friendly to manage several chunks and align them with refference points. Specially because you can mark these on the model (I don't think you can do that in RC). I have not tested georefferenced photos on any of them (it is only included for proffesional packs).

Conclussion: I think I will keep on using RC when I need to produce many models from many photos. But I expect they improve these questions. I would appreciate other oppinions about this softwarre, anyway... My 3 months license has just expired, I have to decide about renewal. :confused:


(Nebulousflynn) #25

thanks again @banalesmuseovirtual

There is actually a thread all about Reality Capture pros and cons here:


(Banalesmuseovirtual) #26

I didn't see it, thanks! :wink:


(Nebulousflynn) #27

PhotoScan Vs. ContextCapture by Decamegas

Posted over on the Acute3D forums, Sketchfaber ssh4 compares the two tools.

Lots of great comparison images:
https://community.acute3d.com/forum/agisoft-photoscan-acute3d-smart3dcapture


(Vlad) #28

No, that is not my comparison. Just found this post on Acute3D forum.

But there is my:
Photos by @geoffreymarchal Fujifilm M-X1 Not a high end camera, but better then my own iPhone6S plus (still enough for "archive quality")

Acute3D/Bentley (My favourite for this moment)


Agisoft Photoscan (this app made me cry :frowning: )

Still fight with OpenMVS, it probably can beat Photoscan in mesh quality.
And want check latest Autodesk Remake (memento) in offline mode with highest settings.

So may be upload new results soon.


(Vlad) #29

Finally made some research yesterday about underlying algorithms used in photogrammetry software. Most research papers I found from OpenMVS modules wiki (and that why i love open-source, because it share such knowledge).

And here some of my IMHO.

If feature detection, camera registration, tie point calculation are mostly have no problem for all modern software, and only vary in some additional feature like manual control/tie points, constrains, etc. So most results are the same and have mathematical representation of quality like reproduction error, etc.

But dense cloud and mesh reconstruction varied too much. If in some application final result is dense cloud, and mesh reconstruction is based on this dense cloud. And this how Agisoft Photoscan work!

Another is based on "High Accuracy and Visibility-Consistent Dense Multiview Stereo HH. Vu et al. 2012." and looks like used in ContextCapture, RC, Memento and OpenMVS in Mesh Refinement step.

If mesh from dense cloud can have estimate polygon count (what we have in Photoscan) and use some kind of poison, VCR or similar mesh reconstruction algorithm.
But "Mesh Refinement" have only some kind of estimated quality settings.

Interesting conclusion:
Only Agisoft Photoscan required masks! Because without masks it produce noise in dense reconstruction step from weak image area (all this weird noise around main object cloud)

CC,RC,Memento,OpenMVS must work better without masks, because can use area outside main object for estimate camera position. Masks are required only if we move and rotate object.
This software need only bounding box, for limit Cloud/Mesh calculation area.
And only in final texture generation step masks are welcome for remove some low quality area in images.

If you better know how all this work. Any new opinion is welcome.


(Vlad) #30

Another one comparison test.

Context Capture ultra precision (still #1)
no masks

Autodesk Remake (ex Memento) in offline mode, highest settings.
no masks

Agisoft Photoscan High, agressive filtration
masks auto generated after first draft 3D model, but still produce too much shit around model

OpenMVG->OpenMVS
without masks.
found that better not use VisualSFM because it did not provide fine tuning. And as result OpenMVG feature detection and camera alignment produce scene.mvs file that as result have better and cleaner result in OpenMVS.

Also tried to test RealityCapture with same sources.
Fast but not as light. Beta. And because of this no mesh view only "capture".
Mesh quality is lower than Context Capture.



(Nebulousflynn) #31

Wow! Thanks for sharing @ssh4 - very comprehensive and some excellent results :thumbsup: Thanks for your previous post about your research into how the different programs too.

At the risk of asking for even more of your time :grin:: Could you share your hardware set up as well?

Any information about your OS, RAM, processor speed might be helpful for other people/organisations wanting to try these bits of software out...


(Vlad) #32

I thought I have new and fast computer when upgrade this spring my old one to new computer with:
z170 chipset matherboard
i7-6700 3.4Ghz
32Gb DDR4 RAM
GTX960 2Gb
Windows 10.

Until I found interesting tool Agisoft Photoscan. That tell me my 32Gb ram and video card is shit.
After this I start searching for alternatives. And now think about Agisoft the same :wink:

But, for serious photogrammetry I think 32Gb RAM And GTX1060 is a start configuration.


(Otto Bagi) #33

I can only agree. The only thing I would add to it that MP actually do matter, more megapixels mean more information the photogrammetry software can work with. I had problems with the alignment of pictures taken with my old Canon 40D, which is fitted with a 10.1 MP sensor, in some specific circumstances especially while working on small flat objects (e.g. coins). Photos of the same object photographed from the same angles but with a 650D (18MP) aligned without a problem.
Another very important factor is the lens, especially when it comes to small objects. A true macro lens is a must, with a standard kit lens it would be impossible to get close enough to frame the object properly. I usually work with a 50mm macro and use a 100mm macro in special cases, they both do a great job but obviously there will be more issues with DoF using the 100mm.


(Otto Bagi) #34

Some of you might find this interesting:


Scaning shining and metal artifacts
(Nebulousflynn) #35

Thanks for sharing @otto_bagi - very interesting!

Really good to meet you today :smiley:

I've also cross posted this to this thread about scanning shiny objects: