Messages

16

General / Clarification on feature, GPS in standard Photoscan?

« on: July 19, 2015, 04:08:06 PM »

Hi all,

I've been looking high and low on the website and documentation (and 'searching' the manual), but I can't find out if the standard version of Photoscan uses GPS EXIF data to help in aligning the image inputs.

Is this an automatic feature, or something I have to set specifically, or not even in the standard version.


I've seen it mentioned before and I was certain it was a standard feature but perhaps I'm just misremembering?



Clarification on the version features page might be useful for others if it is indeed not present in the standard version.


Thanks

Dave

17

General / Re: 1080p and 4K Video --> Point Cloud --> Mesh

« on: March 21, 2015, 07:31:07 PM »

I recently received some feedback on the project I worked on which I noted earlier.

I can't go too far into details probably because it's all NDA stuff, but needless to say, a random rolling shutter video off Vimeo of the site gave me an accurate enough point cloud to generate my project with.

The real life data collected for their intended purpose matches what I provided for them to utilise in a simulated environment which is very reassuring from my perspective.

I'll have to give some rough numbers for it to make sense, but I'd guesstimate at a drone flying about 75m high with the camera looking forward about 35deg-45deg along the direction of travel.
Over about 200 images from the video, and about 2km, I had very little vertical or lateral drift... to the point that I went with the point cloud data for xy and z, over the formal xy plans which were obviously vector based, based off aerial imagery, but with the issue of height parallax distorting features that moved up/down in altitude over the location.



In the end I dumped every 30-40 frames iirc, so I'd go as far as saying something like a Ricoh GR on 1s interval without the rolling shutter issues, and a higher resolution and quality outputs, would result in superior quality results by far.



I just wonder if the DJI Phantom frame could work nicely with a Ricoh GR. Hmmm.


Cheers

Dave

18

Feature Requests / Re: Select front view picture for X/Y/Z orientation

« on: March 10, 2015, 07:51:20 PM »

I'd like to see this feature too.

No matter what you're doing it's always easy to take a few photos aligned to some reference plane that is useful to you, and then use them for alignment.

Quite often you might just drop the camera onto a tripod for example and take a 'flat' shot purely for quick alignment purposes right inside Photoscan, rather than faffing in Cloud Compare or whatever else.


Trying to align by eye with the mouse in Photoscan on the other hand, is quite a slow/inaccurate process.


Dave

19

General / Re: 1080p and 4K Video --> Point Cloud --> Mesh

« on: March 10, 2015, 07:17:31 PM »

I've been capturing images from GE (automated fly-overs with interval capture using FRAPS or whatever), or manual ones, and rebuilding.

The results can be quite good for some quick/dirty projects, vs having nothing at all.

I'm working on a project right now and my rough out source is a point cloud from GE captures.

Infact the only stumbling block is getting a nice alignment after construction. Oh for an align to camera option in Photoscan. I mentioned it over a year ago and it's sad to still not see it added.
So far I need to export camera positions then do a transform in Cloud Compare to the camera position. A bit of a faff!






I've also built a few clouds from random videos on the internet. I worked on one project where someone was nearby a rapidly changing environment that had changed a lot since a site visit, and they'd been flying over with UAV.

Even with rolling shutter issues and so on, I managed to get a really nice rebuild. The point cloud was quite noisy, but with filtering and so on I had a fantastic top-down layout, and the height data was good enough to pick ok average values from.




Agisoft Photoscan is really excellent imo.


OK for some tasks where you want perfection, then perfect inputs need to be captured and then that costs time/money... but for lots of projects it's amazing how you can trim time/costs in capture and get a good technique and still get pretty good results for your needs!



So thumbs up to Agisoft!

But please add an align cloud to camera option!


Dave

20

General / Re: Photoscan non-pro, alignment of cloud methods?

« on: June 13, 2014, 06:07:31 PM »

Well currently I'm loading in Max with Clouds2Max, correcting the alignment, copying the scale/transform/rotation matrix data to CloudCompare, then exporting again.

At least that way I can do things by the numbers, draw rules and levels and double check everything measures up flat/right dimensions etc.


I'll keep having a play with the alignment. I was finding everything was upside down too, and then it struck me it was upside down opposite the last camera reconstructed from... perhaps?!

I'll let you know if I find a consistent way to get alignment squared off to at least one camera... which I'm certain should be possible.


Doing it manually is 'ok', but I still need to fine tune to a level that is really tough without rotation spinners and guides.

21

General / Photoscan non-pro, alignment of cloud methods?

« on: June 12, 2014, 07:37:45 PM »

I was certain in testing that if I aligned my initial camera image (first one in the imported images, or number 1) to be flat that I could then use that as a guide for rotations or corrections of the cloud to be completely accurate.

Ie, in 3DS Max just use the point cloud and rotate the point cloud object say 90deg or whatever to get it aligned properly.

I'm certain I did this with about 3 test reconstructions and assumed it was a good working feature.




But now I'm really struggling to get this behaviour. Quite often it's just one of the random cameras in the set that the scene seems to be aligned with, and it's not easy to get the numerical transformation of that camera from the know camera in the scene.

Using the manual scene rotation feature is just not very useful at all as it's not accurate enough.


Is it possible, or worth adding as a feature, so you can just choose a camera in the list and align the cloud to that camera? You could then easily adjust as necessary with simple transforms (90deg) rather than arbitrary adjustments.


I'll keep having a test as I'm certain this was working before, with camera 1 always being square to the reconstruction.... hmmmmm.


Thanks for any help/advice/thoughts!


Dave

22

General / Re: About Spherical panoramic photos model

« on: May 16, 2014, 10:22:34 PM »

What resolution are the panoramic images?

In the example image too, the distance between the first and last panorama seem relatively close given the size/scale of the environment you want to have a dense cloud constructed for.
Remember a point in space can only be created where it can be seen from several locations. So you tend to want more spaced out with good overlap images.
Right now you have loads of overlap but not much actual movement of the camera.

23

General / Re: Workflow for VFX shots with Nuke regarding camera calibration

« on: May 15, 2014, 05:05:03 PM »

Well if you were to use "camera X" and un-distort in Nuke for any old project Agisoft aside, some kind of process must be available to do that already any way (which we know it is)

Given Agisoft's purpose is to produce a 3D model only, not detect and generate lens distortion profiles for use externally for other purposes, then it's not a surprise to see Agisoft not supporting lens profile generation for all the softwares out there, especially if there are no conventions on storing that data.



Using Photoscan for this kind of work is perfectly reasonable without the camera distortion data isn't it?

Just like ANY 3D asset that you might comp into un-distorted footage, the un-distortion is up to you, not the person/tool that helps you make the 3D asset to begin with?


Or am I missing something? I might be missing something hehe, but I can't see what the problem is right now.
Just un-distort however you would doing any old 3D comped into real life footage?


Hmmm

Dave

24

General / Re: Block Edge Snap issues

« on: May 13, 2014, 05:50:20 PM »

I think you need some overlap, then you need to manage each overlap manually (or via more scripting in other 3D programs)

You'll never get perfectly meshed edges because I assume at the dense cloud edges the geometry is being linearly interpolated beyond the last points.
If the actual points between chunk boundaries don't actually linearly interpolate then each mesh will be reconstructed inaccurately.


The ideal solution would be for edges of chunks to have the mesh creation use some overlap point data (maybe a percentage threshold of the chunk size) for detecting the flow, slice the generated mesh overlap exactly along the boundaries of the chunks, and then they will likely match ok.


Given the benchmarking thread suggests you need huge power for these geometry steps, it'd make sense to make the meshing automatically 'chunk' with overlap, and then re-weld boundaries.
I understand why you need to process continuously to get an appropriate surface reconstruction, but if you use chunks with overlapped processing and cut them then there should be very little risk of different chunks calculating different surfaces over boundaries to not have them meet up (less risk with greater overlap %)


Thanks

Dave

25

General / Re: Workflow for VFX shots with Nuke regarding camera calibration

« on: May 11, 2014, 05:54:10 PM »

There is camera calibration information that Photoscan will create for your images automatically, or there is Agisoft Lens which can generate (probably) more accurate lens distortion information for your camera.

I'm assuming that with this data you can script some distortion parameters in Nuke. Or perhaps the data is even Nuke compatible as there are XML exports available (export cameras)... I assume these fit some general lens distortion conventions that Nuke will be compatible with?

26

General / Re: Workflow - Photoscan to game asset

« on: May 11, 2014, 04:14:41 PM »

Have to agree with Andrew here, Photoscan, like Mocap/Facial rig isnt ever going to be magic bullet, solve all problems and spit out game ready assets.
However, it can produce at least in case of characters, good base with proportions, colors and fairly detail geometry (pending on your setup). If you are working on realistic (believable) game where you need real world architecture, characters, vehicles... Photoscan is the way to go and blows 123d big time.
Making game ready assets is something that needs time and work no matter are you using scans/sculpting or what ever approach.
From my experience in gamedev, finding artists to do retopo/bake/mesh lods.... is much easier than finding artists to do sculpts in quality you would get from 3d scanning and time needed for person to do sculpt/texture is much higher than to do actual scanning.

Admir

I completely agree. You buy Photoscan for the SFM, and it's outputs.

Going too far down the routes for making assets you'd use straight out of Photoscan would seem like a waste of the devs time.

For rapid prototype work the basic tools are already good enough, and for re-working to high quality assets the meshers and tools are already pretty good too.

Not to say there isn't room for improvement, but I hope the developers don't spend too much time trying to make magic bullets, and instead focus on really useful usable SFM features that allow you to get more from your input images to begin with!



In the end nothing is fire and forget, and get good results out at the end. Go laser scanning with ?30,000 scanners, or GPS based helical scanning, and you can suffer from problems still.

Every approach has its pros and cons, and all have problems no matter how good the equipment or conditions are.

SFM and the Agisoft software are excellent in my view! So much so that I've just bought the software. Now for a UAV haha. Maybe a large mono-pod will do for now!


Thanks

Dave