Hi again,
So my question is... it is REALLY possible to classify a point cloud made not by a LiDAR laser scan but through photogrammetry ? I tried but some subsamples may loose this "dimensionality", causing some errors around trees, bushes and vegetation - even if green areas are corrected classified as vegetation - I think that the method of the construction of a point cloud by photogrammetry may be and inconvenient in the classification process, but still not sure of this sentence.
am I thinking wrong or it's just a bad model that I used for this purpose?
qCANUPO (classifier files, etc.)
Re: qCANUPO (classifier files, etc.)
You should ask Dimitri Lague directly but there's no such limitation in the CANUPO algorithm (up to my knowledge).
The main issue may be:
- a too low density of the cloud
- a too high smoothness of the cloud (and that's generally one of the biggest issue with photogrammetry clouds)
Can you show how your cloud looks like whithout colors and with the EDL shader for instance? (you may have to increase the size of the points)
The main issue may be:
- a too low density of the cloud
- a too high smoothness of the cloud (and that's generally one of the biggest issue with photogrammetry clouds)
Can you show how your cloud looks like whithout colors and with the EDL shader for instance? (you may have to increase the size of the points)
Daniel, CloudCompare admin
-
BrunaGarcia
- Posts: 21
- Joined: Wed Feb 19, 2014 10:28 am
Re: qCANUPO (classifier files, etc.)
Yes, of course.
Here it is....
Here it is....
- Attachments
-
- PC_EDLshader.jpg (151.69 KiB) Viewed 194553 times
Re: qCANUPO (classifier files, etc.)
Hum, once again I'm not an expert, but this cloud looks quite "smooth" ('lumpy' may be a better word ;). And on the contrary it's quite noisy/rough at very small scales (the EDL shader is very good to spot the noise).
However the large scales (~5 and more) should still be quite discriminating. But it's hard to 'capture' the boundaries with such large scales. And the smaller scales won't help you much as the cloud is a bit noisy...
Maybe Dimitri would have a trick in this kind of situation?
However the large scales (~5 and more) should still be quite discriminating. But it's hard to 'capture' the boundaries with such large scales. And the smaller scales won't help you much as the cloud is a bit noisy...
Maybe Dimitri would have a trick in this kind of situation?
Daniel, CloudCompare admin
Re: qCANUPO (classifier files, etc.)
Hi Bruno,
answer before I saw the picture:
the downside of use point clouds from SFM is that you have little penetration into the canopy of vegetation. With lidar, even single echo like TLS, you always have points in the vegetation, such that vegetation appears really 3D at relatively small scales. With SFM, vegetation appears 3D at larger scales (i.e. , a bumpy surface) which translates into a classifier of lower resolution, and much more difficulties to correctly identify ground surfaces. I did a test a few months ago on exactly the same scene measured with ALS, TLS and SFM, and the classification of vegetation with SFM was much more difficult, resulting (obviously) in big patches of ground missing compared to ALS or TLS. But it was still feasible with qCANUPO. You just need to adapt the range of scales.
After I saw the picture : ok,that's going to be very tricky. Roughly, you'll be able to tell apart flat areas from bumpys areas but not much more, because even a trained human would have trouble distinguish vegetation from ground (and a trained human eye is still better that qCANUPO ;-). Moreover, it seems that you're working on a hilly area with sharp edges which may pose some trouble too. My best bet would be to look at the RGB field to see if you can use it. Recall that there's actually no groung points below the vegetation, so reconstructing a topography from such sparse data will results in largely uncontrolled interpolation. SFM is great, but not for vegetated area...there only TLS or ALS can really see through vegetation.
answer before I saw the picture:
the downside of use point clouds from SFM is that you have little penetration into the canopy of vegetation. With lidar, even single echo like TLS, you always have points in the vegetation, such that vegetation appears really 3D at relatively small scales. With SFM, vegetation appears 3D at larger scales (i.e. , a bumpy surface) which translates into a classifier of lower resolution, and much more difficulties to correctly identify ground surfaces. I did a test a few months ago on exactly the same scene measured with ALS, TLS and SFM, and the classification of vegetation with SFM was much more difficult, resulting (obviously) in big patches of ground missing compared to ALS or TLS. But it was still feasible with qCANUPO. You just need to adapt the range of scales.
After I saw the picture : ok,that's going to be very tricky. Roughly, you'll be able to tell apart flat areas from bumpys areas but not much more, because even a trained human would have trouble distinguish vegetation from ground (and a trained human eye is still better that qCANUPO ;-). Moreover, it seems that you're working on a hilly area with sharp edges which may pose some trouble too. My best bet would be to look at the RGB field to see if you can use it. Recall that there's actually no groung points below the vegetation, so reconstructing a topography from such sparse data will results in largely uncontrolled interpolation. SFM is great, but not for vegetated area...there only TLS or ALS can really see through vegetation.
-
BrunaGarcia
- Posts: 21
- Joined: Wed Feb 19, 2014 10:28 am
Re: qCANUPO (classifier files, etc.)
Thanks for your guessing Daniel, you were right after all! :)
And Dimitri - even if you called me BrunO...it's BrunA, but its okay, no problem :) - thanks so much for your explanation, I was thinking that it could be a problem with the model and its density, and I'm now totally agree with you that is all about the SFM reconstruction.
As you said, we can do this caracterisation with point clouds generated by photogrammetry, but in fact the vegetation doesn't look like a 3D thing in small scales - what made me choose bigger scales and what took long time for calculation.
I thought it was strange because I trained all week with Otira Gorge + Mont Saint Michel Tidal and it worked very well.
Thanks again for your advice and for this plugin.
I hope I'll have a chance to play with it soon, as long as have better quality point clouds took by LiDAR :)
BrunA.
And Dimitri - even if you called me BrunO...it's BrunA, but its okay, no problem :) - thanks so much for your explanation, I was thinking that it could be a problem with the model and its density, and I'm now totally agree with you that is all about the SFM reconstruction.
As you said, we can do this caracterisation with point clouds generated by photogrammetry, but in fact the vegetation doesn't look like a 3D thing in small scales - what made me choose bigger scales and what took long time for calculation.
I thought it was strange because I trained all week with Otira Gorge + Mont Saint Michel Tidal and it worked very well.
Thanks again for your advice and for this plugin.
I hope I'll have a chance to play with it soon, as long as have better quality point clouds took by LiDAR :)
BrunA.
-
BrunaGarcia
- Posts: 21
- Joined: Wed Feb 19, 2014 10:28 am
Re: qCANUPO (classifier files, etc.)
Forgot to say: In the other hand, I opened a topic asking for a "color classification". I doesn't work, because some parts of the cliff and some bushes that have the same RGB color (same values to the Green channel). So, deleting this value from the Green channel will also delete some parts of the cliff. So no, it doesn't work for this model at all.
Last edited by BrunaGarcia on Tue Sep 29, 2015 8:24 am, edited 1 time in total.
Re: qCANUPO (classifier files, etc.)
Sorry for Bruna ;-)
RGB classification: it's very rare to be able to use it due to shadow effects (especially when you have... vegetation ;-)).
Now, maybe by going back to the actual dense point cloud reconstruction phase and trying to be less strict on the depth processing this could help. My idea: you might actually have more noisy points on vegetation than ground, and this could be a feature that qCANUPO could use.
Good luck
Dimitri
RGB classification: it's very rare to be able to use it due to shadow effects (especially when you have... vegetation ;-)).
Now, maybe by going back to the actual dense point cloud reconstruction phase and trying to be less strict on the depth processing this could help. My idea: you might actually have more noisy points on vegetation than ground, and this could be a feature that qCANUPO could use.
Good luck
Dimitri
-
paulwright
- Posts: 5
- Joined: Fri Nov 06, 2015 3:31 pm
Re: qCANUPO (classifier files, etc.)
All
Just started playing with CANUPO and CloudCompare, and I'm really impressed. Even using the classifiers on the CANUPO website, in a different context (a park scene) I managed to remove trees and bushes, lamposts etc.
However, I'm now moving to a more difficult context. I'm trying to establish a bare earth model of a geomorphological feature covered in a dense vegetation ( a heather covered heathland). I'm assuming that, whilst CANUPO can classify the vegetation, filtering it out after classification will leave me with a lot of 'gaps' in the point cloud, such that establishing a good bare earth model would be impossible. Is this assumption correct?
Has anyone else tried to get BEMs from densely vegetated terrain? if so, what was your solution? can you only revert to expensive software?
Thanks
Paul
Just started playing with CANUPO and CloudCompare, and I'm really impressed. Even using the classifiers on the CANUPO website, in a different context (a park scene) I managed to remove trees and bushes, lamposts etc.
However, I'm now moving to a more difficult context. I'm trying to establish a bare earth model of a geomorphological feature covered in a dense vegetation ( a heather covered heathland). I'm assuming that, whilst CANUPO can classify the vegetation, filtering it out after classification will leave me with a lot of 'gaps' in the point cloud, such that establishing a good bare earth model would be impossible. Is this assumption correct?
Has anyone else tried to get BEMs from densely vegetated terrain? if so, what was your solution? can you only revert to expensive software?
Thanks
Paul
Re: qCANUPO (classifier files, etc.)
Hi Paul,
if you (with your naked eye) are not able to clearly distinguish the ground in your data, then Canupo won't help. The problem with dense vegetation such as grass, is that the laser never really hits the ground (at least for single echo laser scanner, latest generation RIegl instruments might be better). Individual low points are generally the convolution of a bit of vegetation and ground, so you will never be able to isolate the ground, as you don't have it properly recorded in the data.
The best approach in that case would probably to extract the lowest points through the rasterize function.
Alternatively, post a picture of your scene and I'll tell you if CANUPO can help you !
Regards
Dimitri
if you (with your naked eye) are not able to clearly distinguish the ground in your data, then Canupo won't help. The problem with dense vegetation such as grass, is that the laser never really hits the ground (at least for single echo laser scanner, latest generation RIegl instruments might be better). Individual low points are generally the convolution of a bit of vegetation and ground, so you will never be able to isolate the ground, as you don't have it properly recorded in the data.
The best approach in that case would probably to extract the lowest points through the rasterize function.
Alternatively, post a picture of your scene and I'll tell you if CANUPO can help you !
Regards
Dimitri