search

ByeongHyunPak / btc

[CVPR'23] B-spline Texture Coefficients Estimator for Screen Content Image Super-Resolution (Highlight)
MIT License
27 stars 2 forks source link

Qualitative test result is different from that in the paper #3

Open DurYi opened 3 months ago

DurYi commented 3 months ago

Hi there! Thanks a lot to your great work. I downloaded your pretrained model rdn-btc-3rd.pth and tested it with SCI1K dataset by running scripts/test_SCI1K.sh. However, it seems that the PSNR results are slightly different from the results in Table 1 in your CVPR 2023 paper. As shown in the screenshot below, my own test results are marked in green. I wonder whether my test method has anything wrong? Besides, is the pretrained model you uploaded same with the one you used in the paper for qualitative comparison? Thank you. 屏幕截图 2024-07-16 104607

ByeongHyunPak commented 1 month ago

Sorry for the delay and thanks for reaching out!

I have found that I accidentally uploaded the wrong pre-trained weights during the code refactoring process. Unfortunately, I no longer have access to the original pre-trained weights used in the paper, as I deleted my workspace when I left the lab. Therefore, I retrained the model and upload the updated weight. I attach the PSNR results of the re-trained model below. The results show a deviation of less than ±0.10 from the values reported in the paper.

scale 2 3 4 5 6 7 8 9 10
Paper 39.17 34.58 31.10 28.33 26.31 24.47 23.38 22.48 21.89
New 39.14 34.58 31.06 28.35 26.37 24.49 23.39 22.48 21.90
Diff. -0.03 +0.00 -0.04 +0.02 +0.06 +0.02 +0.01 +0.00 +0.01
2 3 4 5 6 7 8 9 10
Paper 34.48 29.56 26.30 24.09 22.69 21.84 21.29 20.90 20.61
New 34.45 29.65 26.36 24.09 22.70 21.85 21.30 20.91 20.61
Diff. -0.03 +0.09 +0.06 +0.00 +0.01 +0.01 +0.01 +0.01 +0.00
2 3 4 5 6 7 8 9 10
Paper 34.91 29.36 24.25 21.57 20.43 19.82 19.45 19.11 18.84
New2 34.94 29.41 24.26 21.56 20.40 19.80 19.43 19.10 18.82
Diff +0.03 +0.05 +0.01 -0.01 -0.03 -0.02 -0.02 -0.01 -0.02