New map engine is struggling with trees/water?

I am making an assumption here. I think Drone Deploy changed the way they process images a few months ago. In some ways its better i.e. Plant health is way better. But here is an issue I have never had until recently.


The lake on left is rendered correctly. The lake on right shows ground from Google Earth

The dark shade in trees is actually Google Earth trees.

Both the images above export giant holes which I guess I am supposed to fill in manually ?

I work in Florida so not capturing water or dense tress is not an option. It appears the new processing engine gives up on complicated stitches? This is not acceptable because the exported image now has giant holes in the image wherever there are trees and water?

Can you please look into this?

1 Like

@WaterWiz water is always going to be a problem because there are no consistent features to get tie points from, but I too have been having issues with trees processed in both terrain and structure mode. Structure was worse. Trees are tough because they can be quite a bit taller than the terrain you took-off from leading to a lack of overlap. They also suffer if there is any wind.

@Kaitlin, this looks similar to the results from my maps that we are troubleshooting.

I would point out that I have over 2 years of imagery under same scenario where this was never an issue until a few months ago?

2 Likes

Exactly. Two years here and I’ve never seen this problem before. I have two maps into engineering right now so we’ll keep this thread updated now that I know it’s not an isolated incident.

1 Like

Hi Aaron,

I’m James from DroneDeploy. I work with our computer vision team on photogrammetry.

You are absolutely correct that we changed our processing algorithm in July to improve our point cloud quality significantly. Part of that change was making is more “certain” of the location of a point in 3D space, and thus more accurate. It also allowed us to better resolve thin structures in maps such as lamp poles, pipelines, and other complex equipment.

The downside as you’ve seen is that, particularly in our “structures” processing mode, we are now more prone to discarding all the points that make up moving surfaces. That includes trees that are moving in the wind, and clear/reflective water with a constantly moving surface.

We will continue to tune this over time, but for now if you are looking for high coverage, I recommend using our “terrain” mode for processing. It includes aggressive hole filling which resolves flat surfaces for lakes, and more faithfully recovers something tree shaped on a windy day.

If this workaround (or the terrain output) is unacceptable for another reason, please let me know and I will bear this in mind as we iterate.

James

What is terrain processing mode? I have never heard of this?

Thanks for addressing this!

When you upload your photos you have the choice of terrain and structure processing. The program will typically default to structure if you have oblique images, otherwise it will choose terrain processing. If you are concerned with ground contour accuracy then you will want to choose terrain mode. If you are trying to make pretty 3D models with vertical elements then use structure mode. If you are trying to model a whole site in structure mode then the flat/horizontal surfaces and homogeneous areas will suffer.

Hi Aaron,

When you click “upload” in a dronedeploy project, you can choose either “Terrain” or “Structures” map type.

image.png

image.png

James

Hi @Jamespipe, et all,

It’s now December 2019, and the same behaviour is still found, with lots of structures missing.

I noticed the trees are the ones who suffer the most, specially on windy days, but I even got blank building walls missing, and to my knowledge walls don’t nudge with wind…

For my line of work trees are one of the most important elements as they are quite frequently the source for terrain and building shadowing which I try to determine, so I need those trees to be there.

It doesn’t matter the detail, just a large body with outside dimensions correct would be fine, we’re producing 3D models from real life scenarios, ghost trees are not in this realm.

Please fix this.

Thanks.

1 Like

The way that image looks tells me that images were not captured from all of the positions needed to reconstruct. From discussing your previous post with you I think that your analytics would be best served generated from the model being processed in terrain mode. This will not give you the pretty sides of buildings, but you will not have any holes in your model and your trees will be solid objects. You can always process both ways, which I do pretty much any time I do a project with a building on it. Each method does things poorly if the imagery doesn’t jive. Keep focusing on even 360 degree coverage like you are painting a house. Not too much detail wasted in any isolated areas. Make your best guess with the available flight plans and start using manual flight to get the tricky areas.

I exclusively use Terrain mode now. No issues anymore with trees or water ghosting or black holes. I suspect if you use terrain with enhanced 3d your issues would be solved.

1 Like

Don’t get me wrong, structure mode is great. The problem is that we are all improving, becoming more expectant and starting to model more and more complex settings. Modeling an office building, church or some other localized structure is easy, but modeling a full site’s content especially with several structures is not.
There is a technology for paving rollers that has a color notification letting you know when an area has been rolled over enough times and passes density. I could someday see a system similar to Live Map that would help you recognize when an area has been covered well enough or not.

:heart_eyes: This would be awesome !

1 Like

Yeah i gave up 3D oblique in mature forest… :grin:(Anyway i don’t need it for now)

1 Like

Obliques should be processed in Structure mode, not terrain. This image looks like it was Terrain processed?

I will say that it takes the absolute right shots to get a good full coverage model. Something has definitely happened to Structure Mode processing as of the last couple months (maybe shorter) that is removing too many small details and causing holes in the models if those perfect shots are not present. I have taken several models that either points were removed or mesh that should have been applied isn’t and reprocessed them in Metashape and there is a tremendous amount of detail. I think this is due to cloud processing and the way they have to decimate the quality of the processing in order to serve the customer load.

@Andrew_Fraser

I processed it in structure mode, shot with crosshatch 3D path, i knew it would be tough for the software, just wanted to give it a shot. I’ll try again this winter around a small patch and this summer without snow.

10-4. I have also had better luck with running the crosshatch, but killing the mission halfway through and restarting with the other direction at a different altitude. I know it is a clumsy workaround, but it is a request for more control over the now native crosshatch including gimbal pitch adjustment. I still use the Auto Flight Modes app for now unless I am doing something simple like a site with a building.

1 Like

I’ll try it next time !

1 Like

For what it’s worth, I too am struggling with trees. Water I can sort of live with, but trees are big part of surveying semi rural village areas with both buildings and trees.

If I select ‘Terrain’ to help with processing the trees, the structures come out rubbish. If I select ‘Structures’, I lose the trees altogether.

Up until a year ago, everything was okay, but something switched. DD changed something in their engine and since then I’ve been unable to have the best of both worlds. It’s a shame as I thought they had just about perfected it, before taking a step backwards.

2 Likes

From what I have seen it looks like what would happen when you are decimating a point cloud to make a DTM. This doesn’t make sense because you should always get the DSM and have the DTM on the backend.

Can you give us some details on this flight? It has got to be one of the worst I have seen.

As I said before, whatever has happened has had an adverse effect unless you capture a billion points. It can work, but it takes too much work. The models are more accurate, but if you don’t capture enough views you get rubbish. The meshes are sharper, but fine details are getting completely removed.

@Jamespipe @Anya, could you give us some guidance on what we are doing wrong? From what I see I know that not enough of the right images are not being captured, but it didn’t use to be and shouldn’t be this hard.