Advice on improving this mosaic? How do you view your stitched images?

[tenboair]

Hi all,

I'm really struggling with trees in my orthomosaics. I'm using a NEX 5T with a 16mm lens for these photos. As you can see in the image, the trees often have sharp edges or are ghosted over other parts.

http://extrazoom.com/image-16882.html?s=huln50x50

This was done with 80/60 overlap, and is two chunks of ~100 photos each @7MB apiece. I generated a sparse cloud, mesh, and texture (since the mosaic doesn't seem to be any better for a dense cloud). There is no GCP data, since I'm having issues geotagging the photos from my flight controller.

Photoscan makes awesome 3D point clouds, but the mosaics are just ok. Is there any way to make a better mosaic? Should I use different software? Better photosets? Better workflow?

Let me know if you would like to look at my photosets or need additional information on my workflow and settings.

Secondly, what software do you use to view your mosaics? Specifically, ones for Macs and PCs and also hosted online. I've found some free ones that work okay, but was wondering about other options.

Thanks for any help!
David

[stihl]

Hi,

As far as I know the ortho image is generated directly from the dense point cloud and that the mesh is just there to keep geometry in place of objects such as building edges and vegetation.

Generating a mesh from the sparse cloud creates a heavily triangulated mesh when compared to a mesh generated from a high or even ultra high dense cloud. This would explain the weird geometry on your trees.

When you say that the mosaic doesn't seem to be any better for the dense cloud, can you explain what you mean by this?

If your computer is up for it try the settings;
Build Dense cloud:
-Quality: Ultra High
-Depth filtering: Mild

and;

Build Mesh:
-Polygon count: High
-Interpolation: Enabled
-Point classes: All

Run those two after the sparse cloud. It yields very nice results for me regarding trees.

[Peter]

Trees usually looks best when making mesh based on sparse cloud, get as flats surface as possible and then export orthophoto

[tenboair]

Thanks! I'll try the ultra high dense cloud. What texture do you use - adaptive orthophoto, orthophoto, generic?

[aggieair]

Good topic.  My trees come out like that as well.  I'll have to try these suggestions.

[StevenF]

I'm new to Photoscan but I've done some other photogrammetry work and I've seen the same problem before. Orthomosaics are commonly made from a bare-earth elevation model (DTM) not a surface model (DSM) which is the initial result of building a mesh from the dense cloud without filtering. If you use the mesh created from the whole dense point cloud then you're more likely to get the sharp breaks created at the edge of trees. A sparse point cloud has less detail and is less likely to have abrupt vertical changes at the edge of a tree so you won't see as many sharp breaks in the ortho.

The real solution is to ground classify your dense point cloud and use a mesh from the ground points when generating your orthomosaic. You can ground classify points in Photoscan using Tools>Dense Cloud > Classify Ground Points. Then build your mesh using only the ground points with Build Mesh > Advanced > Point Classes > Select... and uncheck everything but "Ground".

I've also exported points from Photoscan in LAS format, and then used LASground from LAStools to ground classify the points with some success. You can then generate a mesh in OBJ format from the ground points with las2tin which can be imported back into Photoscan. 

The result will be that only the ground surface and not the trees are ortho-corrected, so you'll see more tree lean. To take care of both tree lean and sharp breaks you would need to be able to create what's called a "True Ortho" which back-fills gaps where the breaks occur using pixels from another image. I'm not sure Photoscan is capable of doing that though.

Also, automated ground filtering of photogrammetric point clouds is a difficult problem. It's not nearly as easy as filtering a lidar point cloud which capable of penetrating the tree canopy. However you can get good results in sparse canopies where objects can be distinctly separated, i.e. the algorithm isn't tricked into thinking the tree canopy is a raised ground surface. 

[tenboair]

Hey Steven,

I tried doing the classify points for dense cloud, and it wouldn't let me build a mesh ("Can't build model"). What are the other settings you used for classifying ground points?

From: Any Class
To: Ground and low point (grayed out)

Max angle: 15 deg
Max distance: 1 meter
Cell size: 50 meters

Thanks for the tips, I have no background in photogrammetry and am just learning as I go.

[StevenF]

I didn't have that problem when working with my current data set but you might want to try looking at the point classification to see the results. You can do that by clicking the "Dense Cloud Classes" icon at the top.

As far as I've found there's no legend or even an explanation of colors for classes anywhere, but you can find out the colors yourself by selecting a group of points and going to tools > assign class... and seeing what the color changes to for different classes. It looks like ground points are brown so if you have a good distribution of ground points then you should be able to generate a mesh (DTM) from them. It's possible the ground classification didn't work well enough to generate a mesh from, but Alexey or one of the more experienced users might know the cause of your error.

I've only recently tried doing point classification in Photoscan so I'm probably not the best person to ask with regard to appropriate settings, but your might want to check out Agisoft's tutorial on the subject: Dense Cloud Classification & DTM Generation with Agisoft PhotoScan Professional http://www.agisoft.ru/tutorials/photoscan/08/

It looks like their ground filtering algorithm is pretty similar to Progressive TIN densification which was I think was originally described in this paper by Axelsson: DEM Generation from Laser Scanner data using Adaptive TIN models http://www.isprs.org/proceedings/XXXIII/congress/part4/111_XXXIII-part4.pdf

The settings you choose will probably be based on the 2d size of the objects you want to remove as well as their height and steepness to the ground. You can probably find what settings work best for your scenario by working with a subset of the data and trying different parameters. I'd also consider searching this forum and the internet in general for recommended settings. I think lasground actually uses a similar algorithm so you might want search the lastools forum too. Good luck.

[tenboair]

Update: Tried Steven's method with a geotagged picture set with 2" GSD and it worked beautifully! Trees have no distortion, and neither does the ground around them. Thanks!

[StevenF]

Hi tenboair,
I just saw this same problem in orthomosaics I'm generating from a set of historical aerial photographs, and I think the issue can also occur in areas with "relatively" smooth meshes/models.

In my case I think what is happening is the photos have poor overlap and the sharp breaks occur on mesh polygon faces in areas near the center of overlap between two adjacent photos. It appears that during the mosaic process Photoscan is choosing what image to map onto each polygon face based on which image has the closest distance from nadir to the polygon. So in areas near the center of overlap photoscan may alternate between two or more images for adjacent polygon faces which causes a blocky appearance in the resulting orthomosaic.

This method of mosaicing is unusual for photographmetry programs and I would guess this is why you aren't capable of obtaining seamlines from photoscan. Many photogrammetry programs draw a single seamline between two overlapping photos during the mosaic process so you may end up with an edge along the seamline, but you don't get these odd blocky polygon artifacts. 

Of course I'm not positive that this is what photoscan is doing behind the scenes. I'm only guessing based on observations I've made so maybe one of the developers can confirm this hypothesis.

Anyway, one way around this problem is to export individual orthophotos and do the mosaicing in another program like ArcMap.  You can export individual orthos by disabling all but one camera and then exporting an ortho.

[tenboair]

Hey Steven,

So you're suggesting orthorectifying the images in Photoscan, but to then mosaic them in a different platform?

That's been my big issue with it as well, but it also reconstructs better since it's not trying to line up different orthoimages like Pix4D does.

I've been thinking about trying out the Arc platform, but am hesitant to dive into such a complex and vast tool. I really just want an easy solution, but there doesn't seem to be one.

[StevenF]

I'm just saying that exporting individual orthophotos and mosaicing them in other software is an option. Currently, it's what you need to do if you want control over seamlines, but I think there are workarounds to seamline control in Photoscan using masks.

I know ArcGIS is capable of mosaicing and seamline editing but I actually haven't tried it yet. I've mainly worked with Socet Set and Erdas Imagine for orthomosacing. I'm sure there are other free programs that are capable of it though. GRASS, QGIS, Orfeo Toolbox, and SAGA GIS are a few possibilities worth looking into.

[nadar]

I have the same kind of problems with orthos of high-rise buildings. Resulting mosaic is horrible.
Finally, I first generate an ortho using photoscan and I use it as a canevas in ArcGIS.
I use the georeferencing tool of ArcGIS and interactively place 20-30 control points on each photo to match their corresponding points in the ortho (selecting only points at ground level). I use the "spline" option to smoothly warp the image. Resulting corrected images are imported in Photoshop using Avenza Geographic Imager plugin and I use the very efficient adjustments layers and control masks of Photoshop to perfectly blend the images in a seamless mosaic.
Takes some time, but the results are far better than anything I get with automatic mosaicing using ArcGIS or ERDAS.