I have trained the CAE model (32x32x32) on ADE20K dataset. Each image will be resized to 1024x1024 and clipped to 8x8 patches. I trained 2 epochs on whole 9276 images. The mse_loss is ~0.00004 for each patch when finish training.
Testing on Kodak dataset, including 24 768x512 images. Results are below:
img_kodim01.png, psnr: 22.4496, msssim: 0.865220
img_kodim02.png, psnr: 26.6289, msssim: 0.839517
img_kodim03.png, psnr: 27.6519, msssim: 0.899806
img_kodim04.png, psnr: 26.6874, msssim: 0.880098
img_kodim05.png, psnr: 21.7232, msssim: 0.883557
img_kodim06.png, psnr: 23.6568, msssim: 0.853647
img_kodim07.png, psnr: 26.5151, msssim: 0.926126
img_kodim08.png, psnr: 20.1571, msssim: 0.891640
img_kodim09.png, psnr: 26.6686, msssim: 0.916273
img_kodim10.png, psnr: 26.6313, msssim: 0.915601
img_kodim11.png, psnr: 24.9720, msssim: 0.886524
img_kodim12.png, psnr: 26.7239, msssim: 0.888142
img_kodim13.png, psnr: 20.5768, msssim: 0.848751
img_kodim14.png, psnr: 23.7810, msssim: 0.874904
img_kodim15.png, psnr: 26.1273, msssim: 0.893920
img_kodim16.png, psnr: 27.1228, msssim: 0.883491
img_kodim17.png, psnr: 27.1712, msssim: 0.928288
img_kodim18.png, psnr: 23.9010, msssim: 0.878718
img_kodim19.png, psnr: 24.2436, msssim: 0.886543
img_kodim20.png, psnr: 25.8101, msssim: 0.916872
img_kodim21.png, psnr: 24.4217, msssim: 0.898863
img_kodim22.png, psnr: 25.8630, msssim: 0.877091
img_kodim23.png, psnr: 26.9843, msssim: 0.897864
img_kodim24.png, psnr: 23.1135, msssim: 0.898676
avg_psnr: 24.9826, avg_msssim: 0.888756
You can download the Kodak dataset and my reconstruction image here.
The grid effect might be the primary loss in the case.
I have trained the CAE model (32x32x32) on ADE20K dataset. Each image will be resized to 1024x1024 and clipped to 8x8 patches. I trained 2 epochs on whole 9276 images. The mse_loss is ~0.00004 for each patch when finish training.
Testing on Kodak dataset, including 24 768x512 images. Results are below:
img_kodim01.png, psnr: 22.4496, msssim: 0.865220
img_kodim02.png, psnr: 26.6289, msssim: 0.839517
img_kodim03.png, psnr: 27.6519, msssim: 0.899806
img_kodim04.png, psnr: 26.6874, msssim: 0.880098
img_kodim05.png, psnr: 21.7232, msssim: 0.883557
img_kodim06.png, psnr: 23.6568, msssim: 0.853647
img_kodim07.png, psnr: 26.5151, msssim: 0.926126
img_kodim08.png, psnr: 20.1571, msssim: 0.891640
img_kodim09.png, psnr: 26.6686, msssim: 0.916273
img_kodim10.png, psnr: 26.6313, msssim: 0.915601
img_kodim11.png, psnr: 24.9720, msssim: 0.886524
img_kodim12.png, psnr: 26.7239, msssim: 0.888142
img_kodim13.png, psnr: 20.5768, msssim: 0.848751
img_kodim14.png, psnr: 23.7810, msssim: 0.874904
img_kodim15.png, psnr: 26.1273, msssim: 0.893920
img_kodim16.png, psnr: 27.1228, msssim: 0.883491
img_kodim17.png, psnr: 27.1712, msssim: 0.928288
img_kodim18.png, psnr: 23.9010, msssim: 0.878718
img_kodim19.png, psnr: 24.2436, msssim: 0.886543
img_kodim20.png, psnr: 25.8101, msssim: 0.916872
img_kodim21.png, psnr: 24.4217, msssim: 0.898863
img_kodim22.png, psnr: 25.8630, msssim: 0.877091
img_kodim23.png, psnr: 26.9843, msssim: 0.897864
img_kodim24.png, psnr: 23.1135, msssim: 0.898676
avg_psnr: 24.9826, avg_msssim: 0.888756
You can download the Kodak dataset and my reconstruction image here.
The grid effect might be the primary loss in the case.