Supplementary material for ICCV submission 667:

Understanding deep features with computer generated imagery

Experiment on ETH-80

1. Quantitative results

We considered the rotations of the different instances as factors of variation and computed the average relative variance explained by the different factors for the different layers of AlexNet:

Rotation Style Residual
pool5 35.4% 21.6% 43.0%
fc6 30.2% 27.7% 42.0%
fc7 29.5% 30.5% 40.0%

The high level conclusions are the same as those of the section 5.2 of the paper, but the differences are less obvious. The variance is explained more and more by the category and less and less by the viewpoint as one progresses in the network. Please see the detailed results for each category here.

2. Qualitative results

We show embeddings for each category of the eth-80 dataset, similar to Figure 2 here. in the paper. Since there were few instances, we show on one side all instances and views and on the other side the average position.