Photoscan vs laser scanner - lidar

[fpbv]

Hi guys

I am getting amazing results using PS, really. But some of my customers are getting somewhat persistent regarding acuracy. They want to know how this compare to high end terrestrial AND aerial laser scanner - lidar. They love the final product we deliver but they want to know about how this will meet their acuracy needs. Since we do ground control surveying, both to terrestrial and aerial purposes and we show the RMS they are fine, but some questions regarding how many points, how dense this is, etc. is getting a little headache.
I would like to have some technical advice on this because I am struggling to find a proper answer. If you have sugestions on links or books it will be nice.
Thanks in advance.

[gEEvEE]

Hi fpvb,

this depends of course on many parameters: the amount and quality of your photographs, the reconstruction method, the scene to be reconstructed, the laser scanner type, the acquired point cloud density of the laser scanner, the incidence scan angle, the georeferencing of the point clouds, the filtering etc. My colleagues and I wrote a paper on comparing laser scanning data of an archaeological excavation (executed in 2008) with photographs processed in PhotoScan. The photographs were not specifically taken for this purpose, but we found the general accuracy to be in the lasers can range, with PS even able to better model edges (which is known to be problematic for most laser scanners). The vertex density was denser than the laser scanning point cloud, but this is of course also scale dependent (it was only a small excavation trench). I hope the paper is published in a few months, so I can link to it here.

Cheers,

Geert

[fpbv]

Hi Geert

Very nice to know about that. I?ve been using this Canon EOS 5D MK II 21.1 megapixel with a 35 mm fixed lens. I already use it to do aerial photos and produce orthophotos with great results. Since I got PS Professional I got some very impressive results, very dense point clouds because I always use 60%-80% overlap between photos, both aerial and terrestrial. I mean, it is a lot of photos but why not?
I am a cartographer engineer here and I always worked with traditional photogrammetric-GIS-GPS systems, but LiDAR I worked once, so this dense point clouds is somewhat new to me, because in the past years I am working as project manager, not so much in technical side of the things.
But I am fascinated about this technology, opens a new ground, new ways to produce maps and such. And PS is just a great piece of software, I can tell you.
By the way, I need more information about filtering, can you give me some more help on this? If you want pm me.
Thanks again.

[Nathan Craig]

Geert,

I'm very much looking forward to reading your paper. Your last publication on PhotoScan was quite good.

Warm Regards,
Nathan

[GPC]

I think the best answer would be to include ground control points (GCPs) as well as quality control points (QCPs) in your ground control network. Use the GCP points in your georeferencing, but not the QCPs. Then, once photoscan performs the bundle adjustment, go through and inspect the vertices located at the QCPs and compare the XYZ values. In this way you can state quantitatively across the project that your sampling of deltas do not exceed .X horizontally, .Z vertically.

BTW, I think this would be an excellent addition to photosynth georeferencing interface.. the ability to input points as quality control points and have photosynth compute an accuracy based on these points. Accuracy is the number 1 question I get from EVERY contractor I talk to, its more important to them than 'how much?'

[gEEvEE]

Hi Nathan,

thanks. Well, it was just a simple paper showing what could be achieved with PS and aerial photographs (the Pro version was not even out at that moment). I will try to post the new papers here once they are out.

As CS states: some accuracy metrics are indeed highly important. It is best to use a fraction of your total amount of ground points for georeferencing the reconstructed model in PhotoScan. The sought after accuracy evaluation can be achieved by treating the remaining GCPs as an independent data set of so-called check points (CPs) to validate the positional accuracy of the computed mesh. In statistical terms, this method is called Hold-Out Validation (HOV). Although the method is known to be sensitive to the spatial distribution of the CPs and even unreliable when only a few CPs are available, the HOV method can be considered a valid accuracy assessment if both the sample of CPs is significant and their location was defined randomly (and covering the whole geographic area of the study site).

@ fpbv: Well, I am currently testing some software packages myself to do some advanced filtering: Meshlab, Geomagic and Rapdiform. When I have some results, I'll try to report on them;

Cheers,

Geert

[fpbv]

@ fpbv: Well, I am currently testing some software packages myself to do some advanced filtering: Meshlab, Geomagic and Rapdiform. When I have some results, I'll try to report on them;

Cheers,

Geert
[/quote]

Hi Geert

Very nice to hear from you. I hope to read your paper soon. Like I said earlier my customers want to know more about acuracy, because the visual aspect is amazing, quite fit their needs.
About filtering I am testing Meshlab, but I think your insights will be nice to everyone. My main concern about PS is some my mesh there a lot BIG holes and some noises on it.

Thanks!

fpbv

[datapolo]

Hi,
with regards to the idea of using only some of the GCPs for rectification I am not sure I agree.

My understanding, which was borne out by a simple test, is that PS only uses the GCPs to scale and orientate the model to a new grid and I assume that it does this in a way that minimises residuals. So given this I would have thought that in the rectification the more GCPs the better to get an orientation that is based on more data - surely? With regards to vertical accuracy I would be more concerned about site related issues such as the length of grass/scrub for example. This can make a significant difference in engineering/development projects.

With good, acceptable residuals that meet the project specs, as an end-user I would not be too concerned about horizontal accuracy if the residuals are OK.

Mike