Supplementary material for ICCV submission 667:

Understanding deep features with computer generated imagery

Experiment on ETH-80: detailed quantitative results

This table presents the detailed results per class on the ETH-80 dataset. The effect visible is the same as the one noted in the paper, but the part associated to the style smaller. This can be in part explained by the small number of instances available and their similar styles.

viewpoint Style Residual
apples
pool5 19.5% 30.6% 49.9%
fc6 19.7% 37.4% 42.9%
fc7 22.2% 38.2% 39.6%
cars
pool5 46.9% 16.4% 36.7%
fc6 40.9% 20.4% 38.7%
fc7 39.9% 22.6% 37.5%
cows
pool5 38.5% 15.6% 45.9%
fc6 30.0% 23.8% 46.2%
fc7 29.6% 26.9% 43.6%
cups
pool5 41.9% 22.4% 35.7%
fc6 39.3% 24.5% 36.1%
fc7 38.8% 26.2% 35.0%
dogs
pool5 35.7% 17.9% 46.4%
fc6 24.9% 27.6% 47.6%
fc7 22.6% 31.8% 45.5%
horses
pool5 41.4% 15.7% 42.9%
fc6 31.4% 23.6% 45.0%
fc7 29.7% 27.5% 42.9%
pears
pool5 27.9% 27.2% 44.9%
fc6 24.4% 33.9% 41.7%
fc7 22.6% 37.8% 39.6%
tomatoes
pool5 31.4% 26.9% 41.8%
fc6 31.2% 30.7% 38.1%
fc7 30.6% 33.3% 36.1%