to reiterate: the goal of the paper is not to
assign semantic meaning to the classes -- this page is intended only for
exploration of the algorithm behaviour. Look not at the labels of the
classemes, but at the images returned on clicking them.
... and image search results may have changed since our training sets were
captured, so should be taken as illustrative only.
the classemes, coming from a general vocabulary, may still contain concepts
or words which may cause offence, or which may retrieve offensive pictures
blue cells have high positive weights, red cells high negative weights
this table is for a 1-vs-all SVM classifier trained on each C256 category,
with binary classemes (329 bytes per image).
The short table in the paper is for 1-vs-all lp-Beta.
this list is sorted by C256 accuracy, from highest to lowest,
but... the caltech-101
subset has been moved to the bottom as the classemes chosen simply exploit the
known deficiencies in the C101 set (despite giving very high performance: 96%
precision@25 for motorbikes, 72% for helicopter, 48% for umbrella).