Created
June 6, 2018 15:57
-
-
Save germank/a542f22be0dad004b18775a7976d1a0b to your computer and use it in GitHub Desktop.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import matplotlib.pyplot as plt | |
# reporting the best results per-team (and top 10) | |
data ={'2011': | |
[0.25770, 0.31010, 0.35960, 0.50450], | |
'2012': | |
[0.15315, 0.26172, 0.26979, 0.27058, 0.29576, 0.33419, 0.34464], | |
'2013': | |
[0.11197, 0.12953, 0.13511, 0.13555, 0.13748, 0.13985, 0.14182, | |
0.14291, 0.15193, 0.15245], | |
'2014': | |
[0.06656, 0.07325, 0.0806, 0.08111, 0.09508, 0.09794, 0.10222, 0.11229, 0.11326, 0.12376], | |
'2015': | |
[0.03567, 0.03581, 0.04581, 0.04873, 0.05034, 0.05477, 0.05858, 0.06314, 0.06482, 0.06828], | |
'2016': | |
[0.02991, 0.03031, 0.03042, 0.03171, 0.03256, 0.03291, 0.03297, 0.03351, 0.03352, 0.03416] | |
} | |
# image net human top 5 error rate | |
human=5.1/100 | |
points = [] | |
for k,v in data.items(): | |
for x in v: | |
points.append((k,x)) | |
x, y = zip(*points) | |
plt.figure(figsize=(9,11)) | |
plt.title('ImageNet competition results', fontsize=22) | |
plt.xlabel('Year', fontsize=20) | |
plt.ylabel('Error rate', fontsize=20) | |
plt.xticks(fontsize=16) | |
plt.yticks(fontsize=16) | |
plt.scatter(x, y, marker='o', facecolors='none', edgecolors='C0',lw=2, s=80, label='Competing systems') | |
plt.plot(data.keys(), [human for _ in range(len(data))], '--', color='grey', lw=2, label='Human performance') | |
plt.legend(fontsize=16) | |
plt.savefig('imagenet-history.svg') |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hi, thanks for the code. Just a quick question, in the ImageNet paper the authors estimate the error rate of an optimistic human annotator (being the correct predicted label by annotator A1 or A2) to be
2.4%
. The5.1%
quoted in this piece of code is referring to annotator A1 only. Annotator A2 (being trained of far less images) exposed an error rate of12.0%
. For reference, see table 10 of the paper: https://link.springer.com/article/10.1007/s11263-015-0816-y/tables/10