search

JihyongOh / DeMFI

[ECCV 2022] Official repository of DeMFI.
81 stars 8 forks source link

Any way to improve results with parameter changes? #7

Open noobtoob4lyfe opened 2 years ago

noobtoob4lyfe commented 2 years ago

The results I'm getting are not of the same quality as the examples shown. Are there any settings or parameters I can tweak to get results closer to the example clips shown?

https://drive.google.com/file/d/1zVtFCj7h2xa-f7cFMpaYtFr3U7N3Qz9y/view?usp=sharing

JihyongOh commented 2 years ago

@noobtoob4lyfe

  1. Your input blurry video has a large resolution with very fast motion, which makes joint problem extreme hard.
  2. As described in Caution of Test_Custom section, state-of-the-art joint methods including DeMFI-Net were trained (also tested) with the predefined blurry formation setting of K = 8 and τ = 5 for Eq. 1. Therefore, the results would not be perfect if the degree of blurriness is different from this setting. Research on making the network to be robust on the degree of exposure variation can be a future work. To improve quality in this perspective, please newly train the network with random blur parameters (uniformly sampling) to be robust towards real-world blurry videos as described in Section V-A-2-f&g of PRF paper ("Video Frame Interpolation and Enhancement via Pyramid Recurrent Framework", TIP2020).
noobtoob4lyfe commented 2 years ago

Thanks very much for the reply! So if we were to describe the ideal amount of input blur in terms of shutter angle, what would be that ideal shutter angle to shoot with? 180 degrees? 90 degrees? 360 degrees? Also, what would the ideal resolution be? Thanks again!