to-do

3d lut oracle
demo competitors (inference one image given path)
eval competitors

results

simplest with penalty 3 channels 12.08 76 val simplest with 7 axis 11.33 76 val (runs/Jun02_17-30-07_weird-power)

hparams: number nodes, axis
training axis or a priori
splines (cubic, simplest, gaussian) 1D
architecture

==================== val clutnet

de76 11.310185 de94 95.77119 mae 0.06297801 psnr 23.209019 mse 0.0074469736 ssim 0.84765625

val curl

de76 11.007149 de94 102.543205 mae 0.06427708 psnr 22.879978 mse 0.007628095 ssim 0.84547865

val ia3dlut

de76 11.2225275 de94 101.27734 mae 0.0648031 psnr 23.0638 mse 0.0077532367 ssim 0.84217143

val maxim

de76 9.745365 de94 88.437935 mae 0.05994275 psnr 23.739414 mse 0.0064541986 ssim 0.8506145

val nse

de76 14.107348 de94 203.34247 mae 0.10562189 psnr 19.259506 mse 0.017525028 ssim 0.8241527

val oracle3dlut

de76 4.298839 de94 19.901468 mae 0.021400109 psnr 31.4657 mse 0.0011626269 ssim 0.90834457

test clutnet

de76 11.319051 de94 108.385414 mae 0.06650369 psnr 22.856966 mse 0.008505996 ssim 0.8229735

test curl

de76 10.689641 de94 97.94769 mae 0.0628198 psnr 22.886576 mse 0.0074530705 ssim 0.8323023

test ia3dlut

de76 11.61062 de94 115.979836 mae 0.07015192 psnr 22.434275 mse 0.009106748 ssim 0.8125029

test maxim

de76 10.15736 de94 93.484375 mae 0.06278963 psnr 23.294592 mse 0.006975092 ssim 0.8241023

test nse

de76 15.490198 de94 236.42274 mae 0.11587683 psnr 18.29282 mse 0.020981716 ssim 0.7902124

test oracle3dlut

de76 4.4504776 de94 20.352753 mae 0.020608082 psnr 31.359543 mse 0.0011194482 ssim 0.8981874

Data

We use MIT 5K with the random250 test split. We create a validation split with the last 250 images of the sorted trainval split. The train dataset is preprocessed by resizing inputs and targets to 448x448 and mapping the input to [0,1], this gives the file processed_train

franchesoni/neuralsplinecolorenhancement