Benchmark report on different architecture and size

[DonaldTsang]

Requesting accuracy, training speed and processing speed (in both seconds and Add-Mult/FLOPS ) benchmarks of different models for Waifu2x e.g. VGG vs ResNet vs Inception (of different sizes)

If you can go the extra mile, compare algorithms with:

• https://github.com/YapengTian/Single-Image-Super-Resolution
• https://github.com/IvoryCandy/super-resolution
• https://github.com/LoSealL/VideoSuperResolution
• https://github.com/IncyLiu/video-super-resolution

[nagadomi]

here it is. https://github.com/nagadomi/waifu2x/blob/dev/appendix/benchmark.md

[DonaldTsang]

@nagadomi so Resnet > Upconv > VGG for all types... I would like to request a few things:

1. A graph for training time vs accuracy/performace
2. Compare Waifu2x to architectures in https://github.com/IvoryCandy/super-resolution that are missing from your benchmark

[nagadomi]

If you want it, do it yourself. The above benchmark describes the results of BSD100 and Urban100, and the results are also described in other papers. Note that waifu2x is not overfitting to fixed bicubic downsampling.

[DonaldTsang]

Here are benchmarks done by somebody else:

• https://github.com/xiaomi-automl/FALSR
• https://github.com/huangzehao/Super-Resolution.Benckmark
• https://github.com/jbhuang0604/SelfExSR
• https://github.com/thstkdgus35/EDSR-PyTorch
• https://github.com/yulunzhang/RCAN
• https://github.com/limbee/NTIRE2017
• https://github.com/yulunzhang/RDN
• https://github.com/Paper99/SRFBN_CVPR19

Here are collections/toolkit collections:

• https://github.com/xinntao/BasicSR and https://github.com/open-mmlab/mmsr
• https://github.com/icpm/super-resolution
• https://github.com/krasserm/super-resolution
• https://github.com/titu1994/Image-Super-Resolution

And here is a list of models:

• https://github.com/ChaofWang/Awesome-Super-Resolution
• https://github.com/YapengTian/Single-Image-Super-Resolution
• https://github.com/LoSealL/VideoSuperResolution