Hi All,
We have become quite interested in autoscoring of images of crystallization
droplets (I think everyone goes through this stage, most grow out of it
eventually) and as part of our work we wanted to develop a well-scored set of
images to use as a training set for machine learning approaches. We decided
that we would try to classify into four classes
Clear
Precipitate
Crystal
Other
This very limited scoring would hopefully be more robust than more granular
scoring, but would give enough information to both help find crystals and to
help find phase boundaries. We have developed a mobile app - CINDER - which
uses swipe technology to assign a score to an image. Every swipe (score) is
then recorded in a database, and we would use the consensus score for each
image as our training set. Four scores also mean that we can assign a
direction to each swipe (swiping right for crystal, up for precipitate etc).
To make the app useful to novices, we have included a training set in the
CINDER app - this training presents the user with an image, then shows the user
how we would have scored it, and gives an explanation of why we scored it the
way we did. Currently our training set ("Cinder Kinder") consists of 75
images. We believe that the Cinder Kinder feature might be quite useful to
train students how to look at crystallization experiments in general.
If you have students or relatives or colleagues who might be interested in a
'citizen science' project, we would be delighted if you could tell them about
CINDER (currently available through Googleplay for android devices, iOS version
coming soon). If you think Cinder Kinder might be useful, and want to include
some of your images in the training set, then we would be happy to do so - what
we would need is the image, and a score (one of the four above) and a reason as
to why you would assign that score. Send any images for inclusion to
[email protected]. Bugs, improvements and comments all welcome!
Cheers, Janet
