Forum

Show Posts

This section allows you to view all posts made by this member. Note that you can only see posts made in areas you currently have access to.


Messages - driftertravel

Pages: 1 [2] 3 4
16
General / Re: estimate image quality
« on: April 27, 2015, 05:04:45 AM »
Image quality can go significantly over 1.0, using an a7r with the sony zeiss 50mm (one of the sharpest lens/body combos on the planet) with the images downsized from 36mp to 12mp shooting at iso 100 f16 I was getting wild quality readings. Of course, it's not really necessary, my canon 100Ds get around .85 and generate similar quality point clouds for 1/6th the price, part of this is the extra depth of field from using a crop lens camera. Still, the sony was nice...

17
General / Re: New workstation specifications
« on: April 17, 2015, 10:01:02 PM »
The one thing with the v2 2687s (though I'm running 2, so there's that) is that they're hot as F**K, I'm using 10 fans in the case and I use the computer to heat my office... especially when running Photoscan, so get some good cooling.

18
General / Re: New workstation specifications
« on: April 17, 2015, 08:20:42 PM »
1. Unless you already bought the 2643v3 and/or the motherboard it's probably worth it, I use two of the V2 2687Ws (which are just as good IMHO as the V3s and you can pick them up used) and it's not slow.
2. PCIE is faster. but the difference is probably not worth the cost, it's much cheaper and basically as fast to run 2 regular ssds in RAID0.
3. I'll defer to dtmcnamara.
4. (or 5.) I've used Win7/Win8/Win10 and it's fast and stable on all of them.

19
General / Multi-Tile UV/Texture Workflow w/keep uvs
« on: April 17, 2015, 08:10:39 PM »
Hello all, I haven't tried this yet, so forgive me if it's a stupid question. I have a rather large project coming up (inside/outside of medium sized house for real-time applications) and, unlike most of what I have been doing, this will require quite a bit more than simply one texture map, as I can only use 4K maps in-engine and my target texel density is quite high. That being the case, I am quite used to using multiple uv spaces on a single model to up the texture quality, however, I'm not quite sure how to project multiple UV maps in photoscan. I know there is an option to make multiple uv maps in the texture generation process, but will this automatically recognize where the uv tiles are on the model when I select "keep uvs"?

The alternatives if "auto-magic" is not an option that I can think of are:
A. to to export an FBX with camera locations into mari and re-project the texture, but seeing as there will be thousands of images I'm not sure how viable that is...
B. Export a vertex colored FBX to ZBrush, reproject onto re-topologized model and use their multi-tile uv feature to bake textures from vertex colors. The problem here is I lose a lot of the texture quality with vertex colors that is the purpose of using photoscan in the first place, as well as the inevitable screw-ups with zbrush re-projection that either take forever or are impossible to clean up...
C. Export many copies of the retopo model with each set of uv's moved to the 0-1 space one at a time and bake textures one at a time... kind of a pain but should work.

Any thoughts?

20
General / Re: Motorized Pano-heads for photoscan?
« on: April 14, 2015, 02:55:02 PM »
Just the kind of advice I was looking for. Thanks!

21
General / Motorized Pano-heads for photoscan?
« on: April 13, 2015, 08:57:16 PM »
Hello, motorized panoheads seem like an easy way to capture images of interiors and such, I'm wondering if this is a good idea from a technical perspective. If I put the panohead in various orientations in a room and took sets of 360degree shots with a 50% or so overlap would photoscan be apt to process these well? I ask because one of the no-no capturing scenarios in the manual is taking photos in stationary 360degree positons...

22
There are programs to re-inject exif data into images processed in photoshop, but I find that just exporting masks from photoshop and keepnig the originals as is works fine and is faster.

23
Feature Requests / Video stripping load in Photoscan
« on: April 08, 2015, 12:42:21 AM »
This might be beyond the scope of photoscan, but I'm finding that stripping frames from 4k video works pretty well, the main problem is that every second or so frame stripped winds up blurry and rates a zero on the image quality analysis in photoscan. If one could load a video clip into photoscan, set a stripping rate (ie. every 10 frames) and during import photoscan could run the image quality analysis on each frame and if it is worthless discard it and take an neighboring frame, it would make taking large scans from video extremely viable, and speed up the process dramatically...

24
General / Re: Which Camera?
« on: March 05, 2015, 04:31:48 AM »
50mm is what photoscan assumes when you input photos with no exif data, so 50mm is kind of the baseline. Realistically the lens you use can (and should) vary depending on what you're going to scan. Lens quality is paramount, fixed lenses generally crush zooms. I use 40mm fixed lens on Canon 100D cameras for medium sized objects, works great and you can get the kit in your price range, comes with a zoom lens too if you want to do different sized stuff, the quality drops off some, but photoscan doesn't have a problem with it, and for game rez textures it's much more than good enough image quality wise. You may see some dropoffs in quality of the sparse point cloud towards the extremes of a telephoto lens, but photoscan is surprisingly versatile. It has much more to do with the way you take the photos than the quality of your hardware in my experience. I compared my $500 kit (100d/40mm) with a $3500 kit of an a7r with a zeiss fixed 50mm lens and the quality of the finished models for game purposes was almost identical. I sold the a7r,  Photoscan had a better time matching points with less images with the 36mp camera with one of the sharpest lenses in the world, but it was totally not worth the price difference imho, unless you have the cash laying around. You can just take a few more pictures with the lower resolution camera to ensure a good alignment. That's my $.02.

25
General / Re: Computer Specifications
« on: March 05, 2015, 04:19:45 AM »
It very much depends on the quality of the point cloud you need to construct, and that's pretty much related to ram. Processor will simply help you do it faster, fast clock speed and multiple cores are important, in that order. If you want high or very high quality point clouds without doing chunks on 1300 photos you're going to want as much ram as you can afford. Seriously. 512gb would not be unreasonable in your situation.

If you want to send me a sample data set I'll see what kind of quality I can get with 128gb ram on my workstation.

26
General / Re: Texturing Problem
« on: March 04, 2015, 06:23:36 AM »
Try a demo of Wrap

http://www.russian3dscanner.com/

See what that does for you. You'll have to read up , but basically use the align function instead of the wrap function. I've had good luck with it, but you're right, the shiny scan isn't helping. One thing you can do is try to clean up the point cloud before making it into a model. Meshlab has some tools for this I think. Otherwise, I've been using cross-polarized light and it works wonders on shiny objects, but it requires special lights and a blackout studio, so it might be extreme for your uses. Check out the difference in the attached photo. Left is just with a polarized lens, right is cross polarized lighting. Plus it's a ninja turtle.

27
General / Re: Texturing Problem
« on: February 26, 2015, 06:45:16 AM »
This is going to be rather hard. There are exactly three ways to go about it I can think of.

1. Make marks on your turntable at the increments you want to take your photos (15 degrees or so). Then take a set of photos at those increments with and without the spray, use the spray textures to build the cloud and then use the un-sprayed photos to build the texture. Photoscan can only build textures based on what you give it, so realistically there is not a good way to do this.

2. Take a set of sprayed and un-sprayed photos, doesn't matter how close the angles are. Build the mesh using the sprayed ones. Build another mesh using the un-sprayed ones, doesn't matter if it's not very good. Use a program like Wrap to align the models. Project the good texture set onto the good model that has been realigned in space to match the right set of images.
 
2. (the easiest way) Just texture the model the old fashioned way using projection painting from a new set of photos in zbrush, mudbox or mari or whatever. This way you don't have to worry about slight differences in the angle of the photos.

28
General / Re: LuxMark 2.0 similiar performance with agisoft in GPU reviews?
« on: February 24, 2015, 07:00:09 AM »
Basic back-of-the-napkin estimations seem to indicate that choosing a graphics card based on LuxMark tests for OpenCL isn't a bad idea. Of course, the guy running 8 Titans is on top of the rankings. If you want the best and don't mind paying for it, wait for the GM200 Maxwell Titan or the R9 390X which should be out later this spring, the R9 390X looks like it might have the edge, but hey, I'm an nvidia guy. Used, I came down on the side of the 780ti for price/performance, but if you want more bang something in the titan range isn't a bad idea. If you have great cooling (especially if you want to run more than 1 gpu) AMD offers some good alternatives. Realistically, at the top end of the gaming level you're probably not going to choose a bad card.

29
General / Scanning Shiny Stuff (Alliteration is Fun)
« on: February 24, 2015, 03:20:37 AM »
So, today I was up for a challenge and wanted to see what I could do scanning objects that are typically difficult if not impossible to scan. I chose my phone, B/C it was in my pocket. The phone is made up of a glass front face, with a little bit of shiny plastic above and below the screen. The sides and back are one piece of machined and polished aluminum.

Overall the items in question is quite reflective, which in my past experiments has left me unable to align photos properly, or given me a mesh with holes, missing chunks, or (in the case of glass) is simply not recognized.

I feel my results were pretty good, I threw everything at it to try and make it come out well. Considering what it is, it could have been worse. Check the attached pics, they show the dense cloud, sparse cloud and renders.

Still, the experiment got me thinking about how people go about using photoscan on reflective objects.  So,I thought I'd post here and see what experience people have with scanning shiny stuff. I'd love to hear tips, tricks and techniques for doing so.

30
General / Re: Masking - is it ever not needed?
« on: February 21, 2015, 10:02:05 AM »
In my experience, when shooting from a fixed position against a uniform background on a turntable masking is the most important. When you rotate around the subject it's nice, but less important. A lot of times in the latter situation the background can actually help align photos, so you mask and then make sure the checkbox for masking in the alignment stage is off. There are easy ways to mask though, especially if you're using a uniform background and a turntable...

Pages: 1 [2] 3 4