We are almost in a position to begin thinking about putting together a methods paper aimed at something like Archaeological Science. It could go somewhere better perhaps, but that is the default.
The idea is to compare different shape description methods and show that Arianna & Stephens Diffeomorphic Shape Matching is better than any others out there.
I have begun to outline the paper here and won't repeat everything, but the basic idea is to at a minimum compare:
- a method based on PCA of the raw shape data + euclidean distance
- a method based on fourier transforms (mocmocs)
- a method based on Diffeomorphic shape matching in non-euclidean space.
There may be others methods, or sub varieties of methods. Momocs implements a few different approaches and perhaps A&S will also come up with various sub methods.
We test these methods in two ways. First, on simulated shapes (Random Walk vases) with the idea that a good method would show a linear increase among RW series with time (as it does for at the variance of a single variable). Second, on a well-curated subset of vases whose classification we know, the idea being that a good method would retrieve that classification.
Here are some questions.
-
Regardless of what we do, should we always begin with registered vases? I think so. That means that we give the data to Arianna, she registers them, and returns registered data to us. We then apply our various methods to the registered vases.
-
Currently I imagine that ``Diffeomorphic shape matching" is one method. But perhaps there are several different sub-methods. If so, then A&S have to think of a few rational choices. Not too many since that will make the paper confusing, but enough to satisfy their own curiosity and make it interesting.
-
Currently I imagine fourier transforms as one method. But momocs implements several methods and it's hard to figure out from the vignette what they are. A&S (S in particular) need to look at this, and help define for me and Han what methods we should try. Han has already had a crack at this, but has used several methods without clearly discriminating among them (Basically he applied the rather terse vignette to the RW data). And other packages may have other methods. We can't do all the methods in the world, but we should at least try a bunch of those likely to be familiar to referees. S will need to advise here.
Stephen: can you look at the momocs vignette and suggest what we should be looking at? Elliptical fourier transforms? Others?
- This paper is certainly not the place for original math. So A&S have to think about what they want to put into here, rather than some mathematical journal. This is really intended for the dumb scientist who has a bunch of shapes and wants to know: what is the best way to analyse them? That will make it really valuable --- even if mathematically boring.
We are almost in a position to begin thinking about putting together a methods paper aimed at something like Archaeological Science. It could go somewhere better perhaps, but that is the default.
The idea is to compare different shape description methods and show that Arianna & Stephens Diffeomorphic Shape Matching is better than any others out there.
I have begun to outline the paper here and won't repeat everything, but the basic idea is to at a minimum compare:
There may be others methods, or sub varieties of methods. Momocs implements a few different approaches and perhaps A&S will also come up with various sub methods.
We test these methods in two ways. First, on simulated shapes (Random Walk vases) with the idea that a good method would show a linear increase among RW series with time (as it does for at the variance of a single variable). Second, on a well-curated subset of vases whose classification we know, the idea being that a good method would retrieve that classification.
Here are some questions.
Regardless of what we do, should we always begin with registered vases? I think so. That means that we give the data to Arianna, she registers them, and returns registered data to us. We then apply our various methods to the registered vases.
Currently I imagine that ``Diffeomorphic shape matching" is one method. But perhaps there are several different sub-methods. If so, then A&S have to think of a few rational choices. Not too many since that will make the paper confusing, but enough to satisfy their own curiosity and make it interesting.
Currently I imagine fourier transforms as one method. But momocs implements several methods and it's hard to figure out from the vignette what they are. A&S (S in particular) need to look at this, and help define for me and Han what methods we should try. Han has already had a crack at this, but has used several methods without clearly discriminating among them (Basically he applied the rather terse vignette to the RW data). And other packages may have other methods. We can't do all the methods in the world, but we should at least try a bunch of those likely to be familiar to referees. S will need to advise here.
Stephen: can you look at the momocs vignette and suggest what we should be looking at? Elliptical fourier transforms? Others?