augustoicaro / sfd-cnn-tl Goto Github PK

Hi @augustoicaro

I'm researching on this topic as well. I just read your paper, the way you process your data looks very interesting. I was also suspicious that the frequency may be one of the reason.
I was wondering do you have the original sgy file of the GSB dataset. Can I access that? I want to export directly from the sgy file to make sure I have the best resolution I can get.
Looking forward to your reply and thank you for your work.

Hi @augustoicaro
I was trying to reproduce your work, and I found that for 5 GSB test example shown in your paper, you only provide 3 mask file (crl2800, inl1796 and inl2011) in this repo. I'm wondering if you could upload the other two crosslines mask file (crl2600 and crl3000)?
Other than the missing mask file. I also found that inl1796.csv is the shape of (498,76) size which is inconsistent with the crossline range (2568,3568 with a step size of 2) and also inconsistent with the inl2011.csv shape (501,76). Could you please help me confirm this?
I tried to reproduce the mask based on your demo in TrainAndSave.ipynb. But the masks I reproduced are not consistent with the example you shown on the paper.
For crl2800, you seem using the haibin's thin version instead of the ASCII file you provided. For inl1798 and inl2011, some faults are missing in the paper.



the most obvious difference is the top area, faults in the paper extend to the very top row while the ASCII file not.
For inl1796, the second left fault length is different compared to the left most fault, also for the rightmost two faults. For inl2011, ASCII file appeared to have more faults on the right side.

I'm not sure why they are not consistent, could you help me confirm these questions? Appreciate your time and reply!