Messages

286

Feature Requests / Re: Photo RAW support

« on: January 09, 2014, 02:12:21 PM »

Windent, RAW files are RAW data, they need to be properly tranformed-baked to usable form, like TIF or JPGs. the RAW file contain all pixels unprocesse, unsharp, flat color, RAW convertor make from this "mess" a usable data......

In my opinion the only real issue with RAW data is the white balance, because the white balance has to be set manually or picked from a color checker card and can vary from photo to photo.
All the other variables can either be done with a single setting (sharpness) or are not so relevant because the images can be imported as 16 bit per channel (exposure, highlight correction, shadow boosting, etc).

Being able to import RAWs directly would be a huge time saver. But I do understand the issues in adding support for this, it would take a lot of effort to support those file formats (and keep them updated and working in the coming years).

287

General / Re: CPU and GPU benchmarks

« on: January 08, 2014, 08:58:38 PM »

Indeed, the number of available cores only applies to the Dense Cloud phase. I can tell for sure because I got it set to use zero cores, and it still Aligns fine.

My comment about the project size was not to say your benchmark is not useful (on the contrary, I love it!), I just wanted to warn you that the results you get from a small project can be slightly skewed because Photoscan is spending a relatively short time on actually crunching numbers.

There are also phases in the Mesh Generation where the process runs entirely on a single core.

288

General / Re: Align photos point limit

« on: January 08, 2014, 08:50:17 PM »

With Medium Accuracy only 20,000 points or less are identified in each photo so changing from 40,000 to 80,000 does not make any difference.

I think the Low/Medium/High is setting how much the images are downscaled before points are extracted (just like the accuracy sets image downscaling during Dense Cloud generation). If you change the setting from Medium to High, the number of points increases quadratically.

We hope to do some experimentation on the max number of points in the near future. We'd really like to find out the most effective setting because it can save considerable time. In the meantime I also use 80.000 or 100.000 points just to be sure. Compared to the Dense Cloud generation the Align phase is still pretty fast anyway.

Calculation time can also differ a lot between projects. For example, I currently have a 300 shot project that takes 6+ hours to align. I think it is because it has trouble finding which photos are matching pairs (because photos are very similar), so it has to match almost all points against eachother.

289

General / Re: CPU and GPU benchmarks

« on: January 08, 2014, 11:36:21 AM »

Agreed, very cool to have Agisoft in a standard set of benchmarks.

I do hope you use a project with a decent photo size. The 'official' benchmark file we use on this forums does not have large enough photos in my opinion. Especially for the Dense Cloud stage, the results in Million Samples/sec deviates way too much with results on real life projects. Because of the small photo size on the sample project, I think too much time is spend on setting up and overhead, and not enough on the actual number crunching.

For example, the AMD R9 290 GPU scores around 750 million samples/sec with this sample project, but on big projects (a set of 300 images with 21MP size) the average score is around 850. And I've processed projects where the same card scores over 1000.

290

Feature Requests / Re: points visibility toggling and point overlay

« on: January 05, 2014, 03:55:54 PM »

+1 for this. It would be a very handy debugging tool.

291

General / How many photos does Agisoft use to create a point

« on: December 21, 2013, 01:56:30 PM »

In the interest of understanding the whole process I was wondering: how many photos does Agisoft use to create a point in the Dense Cloud?

Does it reconstruct points from just 2 photos like Agisoft Stereoscan? So does it work with photo pairs only, or does it reconstruct depth using 3 or more photos?

292

General / Re: AMD RADEON slow response big dataset

« on: December 21, 2013, 01:48:19 PM »

You are running into the memory limit of your graphics card. Not much you can do, except buying a GPU with more memory.

Sometimes it also displays the model, but leaves out every second polygon (so your mesh has tiny holes in a checkerboard pattern).

293

Feature Requests / Re: Lo Minimization or Normal Bilateral filtering

« on: December 21, 2013, 01:44:31 PM »

+1

More mesh decimation options would be excellent!

(although point cloud decimation would still be very useful, please see: http://www.agisoft.ru/forum/index.php?topic=1813.0)

294

General / Re: Can someone help me with this dataset?

« on: December 19, 2013, 11:32:04 PM »

Your images are also very blurry in the corners. The camera data says the pictures were taken with the F3.5 aperture and 1/1600 sec shutterspeed.  You should try to increase your aperture to at least F5.6, ideally higher. When your lens is wide open (F3.5) the image quality will be much less.

Of course with the aperture at F5.6 or higher, your shutter speed will be slower (1/400 sec for example). I don't know how stable your octocopter is, but at a certain point your photos will be blurred because of camera movement. So you have to find a balance between aperture and shutterspeed.

295

General / Re: Hardware

« on: December 19, 2013, 12:02:10 PM »

I have not tested it myself, but more than 2 cards should work fine (if your power supply is strong enough). Windows supports up to 6 GPUs, Linux up to 8. Because Photoscan is using the GPUs as OpenCL devices, you don't need CrossFire / SLI for it to work.

But using more than 2 GPUs is generally not very efficient. The Dense Cloud creation consists of 2 phases: depth reconstruction (GPU accelerated) and depth map filtering (CPU only). In my experience the depth map filtering takes around 40% of the time.  Adding more GPUs will speed up the depth reconstruction, but the depth map filtering will still take the same time.

Lets say the depth reconstruction takes 12 hours on one GPU, and the depth map filtering 4 hours.

1 GPU:
12 + 4 = 16 hours

2 GPUs:
6 + 4 = 10 hours

4 GPUs:
3 + 4 = 7 hours

So doubling the number of GPUs from 2 to 4 will not give you a 200% speedup, more like a 40% speedup. If you have many scans to process you are better off to get 2 machines instead.

If Agisoft would ever add GPU accelerated depth map filtering then this would be a different story ofcourse!

296

General / Re: Hardware

« on: December 16, 2013, 08:14:33 PM »

Yes, Photoscan works fine on these cards. I've got 2 of R9 290 (non-X) in my workstation.

They do between 750 and 900 million samples per second (it varies per project), so they are slightly faster than the Radeons 7970. Price wise the 7970 is still the better choice, it is much cheaper.

297

Feature Requests / Request: Dense Cloud Filtering

« on: December 14, 2013, 10:53:14 PM »

We are struggling with building the mesh for projects with very big Dense Clouds (200 million points or more).

The problem is that the Build Mesh stage always builds the mesh using the entire point cloud. So even if we build a mesh with a 80 million polygon target, it will first build the mesh with the maximum amount of polygons (using all points in the cloud), and only decimate afterwards. This takes a lot of time and we often run out of memory, even with 128GB.

What would be a very good solution, is if we could optimize/decimate the Dense Cloud. For example, in MeshLab you can use "Poisson Disk Sampling", this function filters the point cloud and removes (unnecessary) points.

We would love to have such a function in Photoscan. This would save us a lot of time, and depending on the filtering algorithm it could result in a mesh that is the same quality as the unfiltered Dense Cloud.

298

General / Re: How bandwidth intensive (PCIe) is Photoscan?

« on: December 14, 2013, 12:18:27 AM »

I got my hands on a 1x Pci riser, so I decided to put this to the test by connecting my second GPU through the riser card.

This is with the benchmark project:

Code: [Select]

Device 1 performance: 857.128 million samples/sec (Hawaii)
Device 2 performance: 626.932 million samples/sec (Hawaii)
Total performance: 1484.06 million samples/sec

This is a project with 21MP images:

Code: [Select]

Device 1 performance: 742.359 million samples/sec (Hawaii)
Device 2 performance: 549.72 million samples/sec (Hawaii)
Total performance: 1292.08 million samples/sec
Normally the values are pretty even, so it's clear that running a GPU over PCI 1x connection has a performance penalty (even if it's a PCI gen 3.0 port) If you'd need to use risers then getting PCI 16x versions would be the best.

That said, adding more than 2 GPUs doesn't make much sense as long the depth map filtering phase of the Dense Cloud is not GPU accelerated. Better to have multiple machines with 2 GPUs each.

299

General / Re: Why Quadro s are not good for Agisoft

« on: December 11, 2013, 10:40:57 PM »

Yes, only the Dense Cloud generation uses the GPU. And the second part of this phase (the depth map filtering) is run on the CPU only and also takes quite some time.

So having a fast CPU is very important.

300

General / Re: Why the 3DImage does not appear in the center of the program?

« on: December 09, 2013, 11:19:03 PM »

If you switch to orthographic mode (press '5') the viewport is a bit easier to navigate.