Slave 1 photogrammetry tests

Here's a photogrammetry tutorial I came across. It is palaeontologically oriented, but might still have useful tips.

https://dinosaurpalaeo.files.wordpress.com/2014/08/jopt_01.jpg

Thagomizer, thank you for the link! It sums up the facts that I had to learn the hard way during the last weeks, plus a lot more.

Basically, what I learned until now:
- take as many pictures as you can, from every angle you can. If you think, you already have one from this angle, take another one.
- don't worry too much about perfectionism. A picture with reflections in the glass in front of the object is better than no picture of that angle, and reflections that are present in only one picture can be treated by the software very well.
- If possible, include something with a real measurement in the pictures (I was lucky to have R2 here).
- be honest with the image quality. If you had to use ISO 6400 like me here and noise is high, you can shrink the image size down considerably for the reconstruction. It won't influence the reconstruction in a negative way but improve the runtime a lot.
- If the lighting situation is bad and you have a very dark side, make a color correction before feature detection. I did a reconstruction for the B-Wing, of which the underside was very dark due to the lighting. Using the raw images resulted in the software not being able to align the pictures correctly and it resulted in three individual submodels. With a gamma correction of 1.4 of the pictures of the underside, I was able to use all pictures for the whle model and got a good result.

There is still a lot to learn for me and I am sure that people here can give much better tips, but if I was to start again from scratch, these are the major points to consider.

Cheers,
Thorsten

Polarizing filters would delete half of the light - which would have made it even worse in this particular situation as it had required exposure times of more that 1/10 seconds without tripod (not allowed). As I said, I had settings of about ISO 6400, t=1/20-1/30 and f 4.5. Not a lot margin to reject any photons ;-)

From my experience, noise reduction algorithms only work well for pictures with homogeneous brightness and not on the limit where a lot of the image is almost black with details standing out only marginally. The best algorithm can not bring back information that was never there. This is what I meant with "be honest with the image quality". If there is nothing to see, there is no reason to keep it.

And a lens with f 1.4 will make pictures with such low depth of focus that you won't be able to use them for photogrammetry. For this you need to have the whole object sharp such that feature detection will work well. I found f 4.5 to work well, but you should not work below that.

Thorsten

All of that makes a dang lot of sense ... I´ll give it a try anyway, maybe set up a room with similar lighting to the exhibit and do a test beforehand.

One bit: the digital model is missing the 'stepped' surface on the "nose"(?) forward of the cockpit glazing (speaking about the model being oriented with the engine exhausts/landing pad hull horizontal & facing downwards...)

I suppose you'd seen it already, but just in case... Regards, Robert

Hi,

yes, it's still missing, as it relies on the other three surfaces to be correct first. When they are finished, I'll add the smaller features such as the step you mentioned, the cutouts for the wings and the gun pod fairings.

I had a small step backwards this week as I lost the progress of the week after the last pictures here due to a backup failure (completely my fault). But I have worked on it since then and am nearly finished with those basic surfaces. I'll keep you updated!

Cheers,
Thorsten

Hi, thorst. Sort of a 'bump' as I've starting wading through all the Slave 1 (S-1) goodness here on the RPF because I've just come across the 'big' Hasbro Slave 1 as being a fairly good representation of the ship in 1/48 scale. I've obtained a 'junker' to convert that needs scratchbuilt wings (for a start...) hence my quest to be better educated. It looks like a S-1 to my eye but we all know how that goes; at least the deforms are apparently restricted to the interior and the working bits. I'll keep checking back with you on your progress as I get more into it, think it will be fun to compare the toy to the real deal determination. I will say the toys appears to be symmetrical! ;^P

Regards, Robert

Hi Robert,

I didn't work on this one for some time, as I need to finish other projects first. Plus, I heard that the Star Wars Identities will be in Munic this summer, so I'll go there again and take some shots I need to get a better result in photogrammetry. And perhaps I can accidentally drop a reference ruler in front of it, such that I can determine the scale more precisely. I'll post progress then.

Cheers, Thorsten

Hi,

it's been a long while since the last update, but I needed new data. This Monday I went to Munic to see Star Wars Identities again. This time I was prepared much better for taking pictures for photogrammetry. In 5 hours I took nearly 1900 photos from the Slave 1, the A- and B-Wing, the Star Destroyer, the AT-AT and a few of the Millennium Falcon. The biggest improvement was that I took a 1m-rule with me, which I put in fromt of the cabinets for a few shots of each model, such that the rule can be solved together with the spaceship models to give an absolute scaling object.



And I am amazed how well it worked for the Slave 1! The model changed place and was now inside the large cabinet with the Star Destroyer and most other models. The lighting there is worse, and some of the brighter bulbes are turned on only for about 1 minute every 12 minutes. To get consistent bright lighting, I took pictures only during those intervals and enjoyed the view during the remaining time.

Again I used VisualSFM and CMPMVS for the analysis. This screenshot from VisualSFM depicts all used camera positions (358 in total):



CMPMVS took 13 hours to get this solve:





This was cleaned up in Meshlab and imported into Rhino.

For all who don't want to know dimensional details, please do not read any further.











With the rule in the solve I can now scale the model to real world dimensions.
I can now confirm the following measurements to an accuracy of about +/- 2mm (estimated upper limit of the uncertainty based on the rule's mesh density):

Length of skirt: 588 mm
Width of skirt: 333 mm

To get a feeling of the accuracy of my first attempt, I imported the last iteration of my reconstructed Rhino model into the file, and it matched up perfectly sizewise, with the only larger inaccuracies worth mentioning being in the region where the first photogrammetry solve was bad because of bad coverage. For most other areas of the mesh the distance between my new solve and the first Rhino model is in the order of 1 mm. So I think both the method and the programs used proved that they lead to an accurate reconstruction.

I'll now first focus on processing the other models before I try to make a new model of the Slave 1 in Rhino.

Best regards,
Thorsten

Wow. That's incredible. A lot of work, which seems to have paid off. The final renders look amazing, a bit weird, but amazing all the same.

Excellent! You've got my attention!

Really great job and hopefully if your spot on ( I couldn't resist by the way ) then I'm about 3-4mm out in width and about 6mm in length so not bad at all, really pleased

Make sure you post more as you go and really great work

Great stuff!

Thanks for the comments!

Guy, I'm really looking forward to see your model growing! A real one is always better than a virtual one!

It's still nice to have it though bud and like I said I'll help where I can if you do manage to get it off the screen

I think once the skirt is trimmed its hard to measure untill fixed but I'm slightly narrow on width but that's way better than being over as my last one us at about 348 after its relaxed off the Buck so this time I'm able to rectify a few mm here and there . Ive still to add correct sheet thickness and second skin so this will give me more width , where did you measure 333mm ? At the outer edge of the rear wing well opening?

Guy Cowen said:

It's still nice to have it though bud and like I said I'll help where I can if you do manage to get it off the screen

Click to expand...

I really appreciate it!

Guy Cowen said:

I think once the skirt is trimmed its hard to measure untill fixed but I'm slightly narrow on width but that's way better than being over as my last one us at about 348 after its relaxed off the Buck so this time I'm able to rectify a few mm here and there . Ive still to add correct sheet thickness and second skin so this will give me more width , where did you measure 333mm ? At the outer edge of the rear wing well opening?

Click to expand...

It's the maximum width of the oval without the cutouts, and a first approximation, so give or take a mm (the uncertainty of my first scaling attempt - the +/-2mm as mentioned above - is larger than the difference between the total width of the oval and the distance between the edges you mention, so I didn't yet try to refine it). When I will start the CAD work, I'll first revisit the scaling and try to push it to +/- 1mm, I have Ideas how to do this.

Regarding your skirt pulls: the relation of the difference between your skirt and my measurements in both directions - length and width - seems to be about the same factor as the total length and width (3mm/6mm ~ 588mm/333mm). In principle that should allow you to get almost exactly my skirt model if you just cut the skirt a little bit lower, if you have enough room. I sent you my latest CAD model some time ago, you may be able to check this against your model to see if that can be done. Just don't look at the upper starboard quadrant of the skirt, that was not covered enough with my first photogrammetry attempt.

But of course, feel free not to trust these numbers too much, I don't want to be guilty if it's wrong in the end. I only do my best with the limited tools I have.

Don't worry bud mines set in stone as it all links together re armature, all parts fit where they should and it's full steam ahead , for a few mm I wouldn't amend it at all and could even just pull over a pull on the Buck but would loose the trim lines so just rocking with what I have. 333mm sounds about right as internal is around 326, with 3mm of skirt either side so very close,

I used to fly aerial mapping and spent a ton of time on photogrammetry missions. Never thought of applying it to a model. Smart!

The only thought I didn't see (forgive me if I missed it), was to calculate the distance between the 1mm rule and the model itself. As well as a small factor for camera lens distortion (though at that distance, I it may have very little effect). Between that and measuring a known kit part on the model, I would think you would be assured of being dead on.

Awesome work!

Well I certainly wasn't the first one to use photogrammetry for this task, in fact, I got the idea here from another thread on the Millennium Falcon.

Lens distortion is cared about by CMPMVS, as far as I understand the program. You can also overlay your reconstructed model with the photos using the reconstructed camera positions, and it's dead on. Additionally, lens distortion should only affect the general mesh accuracy but not bias any directions, as long as the coverage of angles around the model is not very limited. In this case, I have nearly 360° in azimuth and +/-30° in elevation, so any lens distortion effect should average out, even if it would not be cared of as a free fitting parameter. The advantage is really that you can determine camera position, angle, focal length and lens distortion plus the model geometry all at once. At least that's how I understand it.

Lens distortion has a huge impact on other kinds of photoreconstrution methods though where only one photo is used, so you are right in principle.