MeanIU of FCN32 lower than the result in Shelhamer et al. (2016)

Question

MeanIU of FCN32 lower than the result in Shelhamer et al. (2016)

yasuru opened this issue 8 years ago · 6 comments

I'm training FCN32. The meanIU on validation data set is ~30 % after 100000 iterations.
It can't reach the result (63.6 %) of Shelhamer et al. (2016). Is this correct?
Would you tell me what is the difference?

Answer 1 · 2017-02-06T10:54:56.000Z

According to paper at http://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf, page 6, the MeanIU is (59.4) for FCN-32s.
The difference is

dataset (pascal2011 segval in their work, 2012 mine)
iteration (maybe they're larger than mine)

Maybe you can try training with voc2011 with iteration larger than 100000.

Answer 2 · 2017-02-06T11:45:31.000Z

Thank you for quick response.
I will try it.
I have another question. I'm training FCN16s using pretraind FCN-32s with meanIU ~30 %. The meanIU of FCN16s is 22 % after 100000 iterations. It is lower than meanIU of FCN-32s. Have you met this problem? I'm using the following parameters.
・learning rate: 1e-10
・momentum: 0.99
・weight decay: 5e-4

Answer 3 · 2017-02-06T12:24:28.000Z

Actually I've never tried training FCN-16s for pascal dataset.

Answer 4 · 2017-02-06T16:31:39.000Z

Noted with thanks.

Answer 5 · 2017-02-08T02:49:02.000Z

@yasuru I met the same problem about The accuracies of FCN16s is lower than FCN-32s. What is the matter?? @wkentaro

Answer 6 · 2017-02-19T16:22:00.000Z

Hi, I think I found the cause of this issue.
That is because I didn't copy the weight of pre-trained model (VGG) to FCN32s regarding of fc layers.
weight of fc layers can be reshaped to use it as the weight of conv layers.