search

xialeiliu / RankIQA

The rep for the RankIQA paper in ICCV 2017
https://xialeiliu.github.io/RankIQA/
MIT License
453 stars 117 forks source link

problem when use in my own dataset #21

Closed sunbinbin1991 closed 6 years ago

sunbinbin1991 commented 6 years ago

Sorry to bother you again. I have ask for your Wechat username several days ago. Maybe you have not notice it. Here is my Wechat username： sunbinbin5210 Here is my email: binbinsun1991@163.com I have some question as below: 1: CROP or RESIZE: I want to train a model used in my project for select high quality image.Actually, my picture almost are small,its not proper for crop a picture have size with 224*224,so I want resize my picture for training.Whether crop image and resize image have a big difference in training? 2: Convergence: I used your waterloo dataset for rankIQA training , and then fine in my own dataset, I find its really hard convergence, and LCC and SROCC is about 40%, is something wrong in my training, also I use MobileNet for training not VGG-16 for speed. 3: DataSet:,I can user distor_generation.m to generate my own dataset, so if I use my own dataset for Rank Training ,and fine -tune in my own dataset, could I have a better result? Looking forward your reply...

xialeiliu commented 6 years ago

Ok, I will contact you on wechat.

  1. I think you could do resize, but you should keep the aspect ratio the same, so you don't distort the content of images
  2. I don't know what your dataset looks like, could be that you dataset is too challenging?
  3. Yes, That would be interesting to try.