Openmax behaves worse than Softmax about osdn

[kevinlemon]

Many thanks !

[aadeshnpn]

Thanks Kevin.

Based on the paper I have implemented the authors report OpenMax performing better than Softmax. My implementation also supports the claim. The paper was from 2014. I heard they have now better methods than Openmax. Take a look at the following page for the researcher

https://www.wjscheirer.com/projects/openset-recognition/

If the code in this repo doesn't work. Let me know.

[kevinlemon]

I am sorry for the late reply. However, In the “Softmax.ipynb” the OpenMax behaves actually worse than the softmax.
For example: In MNIST test. The first test image, the actual label is 0, the openmax predicts 9; the third image, the actual label is 7, the openmax predicts 2; the fifth image, the actual label is 0, the openmax predicts 9; whereas the softmax predicts correctly for all the above images.

[aadeshnpn]

Thanks for pointing that out.
Since we are just picking out random 5 images that the openmax just predicted correctly 3 out of 5 samples. But when I ran it through the whole test set, it accuracy was about 95%.
Look at the Chinese dataset example, how openmax correctly classified those characters as unknown where softmax was giving wrong predictions. This was what openmax was designed to do: detect distributions that are not within the range of training distributions.