qCANUPO (classifier files, etc.)

[daniel]

Well, the plugin do compute this statistics in 'training mode' (when you create your classifier, when you see it's 2D equivalent, there's a button to show all these pieces of information. However that's easy here because the user has provided the right subsets to the trainer.

Once you have applied the classifier... not sure what you can do apart from what you did...

[Dimitri]

Hi Lena,

to evaluate the quality of your classifier you need to apply it on a dataset that you have manually labelled (and not your training dataset). Once you have applied your classifier on all your data, you extract only class1 from the Canupo classification. From this dataset you extract the two classes that you have manually labelled and you'll be able to compute your precentage of success and error 1. You do the same for the class1 predicted by Canupo.

It's a tedious process. If your dataset is not too large you could export it after the qCanupo classif to matlab (or python) to easily extract the percentage of error much faster.

Hope this helps

Dimitri (Lague, from Brodu and Lague ;-))

[Lena]

Hi Daniel and Dimitri,
thank you very much for the fast answers! Fortunately, my dataset is quite small. So the way to export it to matlab will be the best one, I didn't think about something like this - so thanks :)

[murraynairn]

Hi Daniel & Dimitri

I have been using the Canupo plugin for my MSc Project. One of the things I have been asked in review of one of my chapters is exactly what the X & Y axes show on the graph which displays when you train a classifier?

I have attached a picture below so you know exactly what graph I am talking about. There are no X and Y axes labels included. It would be good to have an understanding of exactly what the graph is showing.

Many Thanks,
Murray

[daniel]

If I remember well, the two main axes are obtained by a kind of PCA analysis (https://en.wikipedia.org/wiki/Principal ... t_analysis). The Canupo descriptors are corresponding to vectors with a lot of elements (2*N, where N = number of scales). The classification is therefore done in a 2N dimension space. This would be impossible to display it in 2D, so the plugin projects the vectors along the two main components (obtained by PCA) .

It may be a bit simplified, but that's the spirit ;)

[murraynairn]

Many thanks Daniel.

So am I right reading the graph I posted that there is greater variation in the Principal Component 1 (along the X-Axis) than along Principal Component 2 (along the Y-Axis), hence a more vertical plane of separation, or the other way around?

Furthermore, Once classified, how is the Class Confidence calculated? And I take it the value essentially means the classifier is (e.g.) 95% (0.95) confident that this point is classified correctly?

Thanks for your help.

Murray

[daniel]

The difference between the two first principal components may be very small (as they are the two most important ones), but I guess it's right to say that the first one is bigger than the second one.

Mind that they are internally computed and you don't have access to the Eigen vectors, so it's a bit 'arbitrary' in a way. It's mostly a way to represent things in a user friendly way, and to let the user update the boundary manually (you can add and move the vertices boundaries in this 'plane').

And for the confidence, your interpretation of the meaning is right. For the exact way it is computed, I fear you'll have to dig into the Canupo article ;).

[rocksarecool]

Hi guys, I'm currently completed a project using TLS point cloud data of a sedimentary rock formation and I was hoping to use the CANUPO plugin to attempt to distinguish between the different types of sedimentary rocks. The target area is made up of mudstone and sandstone which are easily distinguishable from photos but somewhat more difficult to determine in a geometry sense. The point cloud is rather dense with about 5mm spacing.

Is this something that the plugin would be capable of doing successfully? I've been specifying training data of around 6,000 points for each class but I'm unsure about what scales to use. Results are average at best so far.
Any help would be greatly appreciated! :)

[matknaak]

Hi,

when you say that your mudstone and sandstone is easily distinguishable by photos, that means they have different colours etc. To my understanding CANUPO plugin classifies based on geometries not further attributes. Maybe try to use the intensity value of your TLS data to discriminate between lithologies - this might be much easier. The only problem there is that you might have to aply a correction for the distance to the scanned object. It wood be off course bloody brilliant if there would be a build in tool in CC to do so, but their ain't...

cheers

mat

[rocksarecool]

Hi,

when you say that your mudstone and sandstone is easily distinguishable by photos, that means they have different colours etc. To my understanding CANUPO plugin classifies based on geometries not further attributes. Maybe try to use the intensity value of your TLS data to discriminate between lithologies - this might be much easier. The only problem there is that you might have to aply a correction for the distance to the scanned object. It wood be off course bloody brilliant if there would be a build in tool in CC to do so, but their ain't...

cheers

mat

This is true about the geometry. RGB doesn't factor in at all. I didn't know this originally which is why I've given it a shot. It's still incredibly useful for cleaning up the rock face of vegetation and what not. Was hopeful that it might be able to pick the differences (sandstone is much smoother whereas the mudstone is much more textured and 'crumbly') but the resolution of the point cloud may not be good enough to pick up such small differences.

Cheers for the tip about intensity I will give it a go.