justinengelmann
commented
1 year ago Hi mystvearn,
Thanks for your interest and for raising this issue.
The svm can be found here: https://github.com/justinengelmann/QuickQual/releases/tag/1.0 or in the sidebar of the main repo under releases.
That's an interesting case, is the image you uploaded that full resolution of what you device acquires? The image shows up as 320x240 on my end, which is quite a bit lower than 512x512 and non-square.
After resizing, the image looks like this:
For that one, I get Predicted p(bad): 0.8483 from QQ-MEME and array([[0.01730302, 0.39482401, 0.58787297]]) from the SVM.
If I crop the image to be square and to remove the black areas (img.crop(box=(35, -5, 285, 245))), it changes to Predicted p(bad): 0.7493 and array([[0.02205677, 0.45418776, 0.52375547]]), so slightly better quality scores. After those transformations, the image looks like this:
To me, this is quite a reasonable score, as the image ends up being very blurry.
If you need to assess many of these images and this is indeed the max resolution your device provides, it would be possible to try a version that was trained on a lower resolution than this, maybe that would work better. Feel free to send me an email if that'd be of interest.
Best wishes, Justin
Hi justinengelmann,
I came up with the article and had to immediately jump onto my PC to try out your code. The example with QuickQualMEME works well despite that the result is not so favorable. I may guess that the not-so-well output is due to the fact that my image comes from an automatic mobile fundus device.
Hence I would like to test the big model but cannot find "quickqual_dn121_512.pkl" in the repo. Could you please upload or let me know how to generate this one?
Rgds.
PS: this is an example of the image I use for testing MEME model