Test other drawing examples
Closed this issue · 9 comments
Currently, our apple example is too homogeneous, i.e., there aren't many interesting differences.
Inspired by http://formafluens.io, please test a couple of other datasets to see if we can find interesting regional clusters of differences. On their website, they highlight these differences as diverging examples. Some promising datasets to test are:
- teapot (the data is already prepared)
- banana
- boomerang
- ice cream
- watermelon
- snowman
For each example please prepare the data using google-quickdraw-embedding.ipynb:
- Load the images
- Compute their similarity
- Compute the umap embedding based on the similarity
- Load the images in piling.js
- Arrange the images by the embedding
- Cluster by overlap (maybe in combination with region) to see if there are any visible differences
- Take a screenshot of what you found and attach :)
The fish swims to different directions in different countries
Power outlet in different countries
Bananas in different countries
Snowman. Two snowballs in some countries and three in others.
Traffic lights. Taiwan and Japan's are horizontal
Watermelon. Slices and the whole one
Power outlet:
In North America, it's obviously two holes:
In East Asia & Pacific, the images are diverging:
In Europe & Central Asia, mostly it's one hole:
Snowman:
In North America, the snowman has 3 balls:
In East Asia & Pacific, many countries' snowmen have 2 balls:
In Europe & Central Asia, it's also 3 balls:
Closing this as we've tested a bunch now :)