Size vs complexity
Thanks for a fascinating project! My background is in marine geophysics, particularly analysis of sonar images.
You mention that you resize the images, because there may be a tendency for humans to classify larger objects as more complex. However, in resizing them you are, presumably, changing the resolution, so that smaller spots, once magnified, will appear fuzzier. Might not this also bias the perception of complexity? Certainly I find that, all else being equal, I tend to see more complexity in the higher resolution images. Perhaps your analysis could compare complexity of a) similar resolution features, and (b) images features of differently sized/resolution images?
Finally, there are some image pairs that generally seem to have similar complexity. How about a third option of 'both the same'?
Hi @RSearle! Thanks for your comments. In fact, the analysis of the results after this changes in resolution have to be treated with special care. In most of the case I would see the higher resolution images with larger complexity as you do, but I have gotten different comments from different people, seeing more complexity in the fuzzy images. We've got already our hands in our first dataset of classifications, let's see what that tells us. We will have to compare them in multiple ways, being these two ways that you mention - (a) and (b) - one of our first steps. We will post all our findings in the blog!!!
Regarding your last point... I'm think that if we allow be mark both as the same we won't be able to distinguish them... but maybe one of the people more familiar with the classification algorithm could tell us. There must be some small difference to make you think one is more complex than other - or we hope so 😃
Great discussion... Yes the size bias is something we have been brainstorming about since the previous round of Sunspotter. One thing we can do is to look at the Sunspotter complexity distributions for different 'Hale Class'. For instance we can see what is the median complexity for Alpha-class sunspot groups compared to Beta-class sunspot groups, compared to 'Gamma-class'. Then we can do the same comparison for different size bins. If there was no bias (and the classes are accurate) then we should see the complexity values as alpha < beta < gamma, regardless of size. We can send around a plot, when one of us gets some time to test that.
As for the 'fuzzy bias', this is something that we can definitely test after we get the classifications in, at the end of this round.
And, as @dvdgc13 mentions, we want people to choose one or the other, since choosing 'the same' unfortunately doesn't give us any ranking information to work with. But, we can tell that people thought two sunspot groups looked the same, because the complexities will be very similar, and the last few clicks classifying each one will not be able to differentiate their scores by much. In fact, what we found in the last round is that on average sunspot groups looked pretty similar to +/-50 of their neighbors.