search

ThediidehT / LKFN

Large Kernel Frenquency-enhanced Network for Efficient Single Image Super-Resolution
MIT License
5 stars 0 forks source link

About the training dataset? #2

Open wbhhello opened 3 hours ago

wbhhello commented 3 hours ago

I truly admire the outstanding work you've done. While comparing models, I noticed that both you and MDRN use the DIV2K dataset along with the first 10K images from LSDIR for training. Most lightweight super-resolution work, however, only uses the DIV2K dataset. Have you had the chance to compare the differences between these two approaches? I eagerly await your response.

ThediidehT commented 3 hours ago

I apologize for the confusion. The model in the repository is actually trained using the DF2K dataset, while the competition version is trained using the LSDIR dataset as described in the paper. The pure DIV2K dataset is relatively small, and current mainstream lightweight models generally use the DF2K dataset.

wbhhello commented 2 hours ago

Thank you for your response. I noticed that the version on arXiv does not mention the DF2K dataset. Could this be due to a typo in the arXiv version? Additionally, were the metrics in Table 4 obtained from training on the DF2K (DIV2K + Flickr2K) dataset? I look forward to your further clarification. Thanks!

ThediidehT commented 1 hour ago

yes,it is actually trained on the DF2K