Astrophotography

CODE: GitHub
DATATYPE: images
SOURCE: Yahoo Flickr Creative Commons 100M Dataset
YEARS: 1994-2014
SIZE: 5,146
SUBSET: "astrophotography"
VIZTYPE: map
FEATURES: tag space
FEAT SRC: yfcc100m autotags
ORIG FEAT DIMS: 731
EMBEDDING: t-SNE
BINS: 129
GRIDDING: fill
CRITERION: adjacency
ANNOTATIONS: tag lists
% ANN: 56
PROJECT: Crockett and Gurevitch (2016), in progress

This is a "map" visualization, meaning that the images are plotted in a coordinate space — here, a 2D t-SNE embedding of a 731-dimensional tag space produced by Yahoo Flickr. Since Yahoo Flickr's classifiers use sophisticated measures of visual similarity, its outputs produce a sophisticated visual similarity space. The images are gridded into 129 bins in this space, with some images allowed to plot adjacently to their true locations in order to increase the plotting volume. The map is also partially annotated with short tag lists. Blue markers point to the annotated images. We use this map as a starting point for our analysis: it gives us the visual landscape of our dataset, and the tag features give us a powerful computational foothold in this landscape. This is part of an in-progress project.