search

LSIbabnikz / DifFIQA

Official repository of the paper "DifFIQA: Face Image Quality Assessment Using Denoising Diffusion Probabilistic Models" in proceedings of IEEE International Joint Conference on Biometrics (IJCB) 2023.
18 stars 2 forks source link

Score of diffqa very high score with both lower quality image and higher quality image, and it difference regression model, the scores deviate a lot. #2

Open ChuRuaNh0 opened 11 months ago

ChuRuaNh0 commented 11 months ago

Can you help me? Thanks a lot.

LSIbabnikz commented 11 months ago

Can you maybe explain a bit more about your situation?

If I understand correctly, you are trying to use DifFIQA (the unsupervised approach using the DDPM model) with different regression models? The results will be highly dependent on the regression model used, so no real surprises there. We have not tested many FR models, but all of the state-of-the-art models presented in the paper should work well at this stage.

In general, the range of values predicted by DifFIQA is not that large, most images will be in the range [0.7, 0.9]. More importantly, the lower quality images should be ranked below the higher quality images.