search

guanghaoyin / CVRKD-IQA

Pytorch code for our AAAI2022 Oral paper "Content-Variant Reference Image Quality Assessment via Knowledge Distillation"
MIT License
56 stars 7 forks source link

A question about the results of HyperIQA #2

Closed chencn2020 closed 2 years ago

chencn2020 commented 2 years ago

Hi, there is something confuses me about the experiment results of HyerIQA in table 1, because I'm just a beginner in IQA. The Srocc and Plcc results on TID2013 and KonIQ-10k are 0.686 0.721, 0.332 0.338 respectively, which is far from the results in original paper. So, did you also retain the HyperIQA with multi-patches instead of following the original experiments settings?

guanghaoyin commented 2 years ago

Yes, our experimental setting is different from HyperIQA. For fair comparisons between FR/NR/NAR methods, we retrain those NR/NAR-IQA SOTAs following the commonly used experimental setting of FR-IQAs. All NR/NAR-IQA SOTAs are retrained on synthetic Kaddid-10K and cross-evaluated on LIVE, CSIQ, TID2013, and KonIQ-10K.

chencn2020 commented 2 years ago

懂了, 多谢(ㅅ´ ˘ `)