Supplementary material for ICCV submission 667:
Understanding deep features with computer generated imagery
Quantitative results for the lightning experiment of section 4.2
We present here the quantitative results for the light variation experiment. They are not used in the paper but the corresponding embedding is presented figure 4. Note that on the contrary to the other factors, the light is not well separated from the style. However, it is also considerably less important as the networks we studied is largely independent from lighting.
Results
car
light
| |
pool5 |
fc6 |
fc7 |
| Factor |
| Places |
0.9% |
1.1% |
1.0% |
<
5 |
4 |
4 |
| AlexNet |
1.2% |
1.6% |
1.7% |
<
5 |
4 |
4 |
| VGG |
1.0% |
1.5% |
1.6% |
<
5 |
4 |
4 |
| Style |
| Places |
75.4% |
75.2% |
76.1% |
<
203 |
76 |
76 |
| AlexNet |
77.4% |
75.1% |
75.1% |
<
181 |
115 |
115 |
| VGG |
73.5% |
72.5% |
73.5% |
<
242 |
113 |
113 |
| Residual |
| Places |
23.6% |
23.7% |
22.8% |
<
888 |
446 |
446 |
| AlexNet |
21.4% |
23.3% |
23.2% |
<
798 |
671 |
671 |
| VGG |
25.5% |
26.0% |
24.9% |
<
1000 |
726 |
726 |
|
chair
light
| |
pool5 |
fc6 |
fc7 |
| Factor |
| Places |
0.2% |
0.2% |
0.2% |
<
5 |
3 |
3 |
| AlexNet |
0.2% |
0.2% |
0.2% |
<
4 |
3 |
3 |
| VGG |
0.1% |
0.2% |
0.2% |
<
5 |
3 |
3 |
| Style |
| Places |
92.0% |
92.9% |
93.4% |
<
500 |
240 |
240 |
| AlexNet |
93.0% |
93.5% |
93.8% |
<
448 |
347 |
347 |
| VGG |
91.8% |
92.7% |
93.0% |
<
604 |
296 |
296 |
| Residual |
| Places |
7.8% |
6.9% |
6.4% |
<
1000 |
1000 |
1000 |
| AlexNet |
6.8% |
6.3% |
6.0% |
<
1000 |
1000 |
1000 |
| VGG |
8.1% |
7.2% |
6.7% |
<
1000 |
1000 |
1000 |
|
|