search

Reyhanehne / CVF-SID_PyTorch

Official implementation of the paper "CVF-SID: Cyclic multi-Variate Function for Self-Supervised Image Denoising by Disentangling Noise from Image" (CVPR 2022)
70 stars 11 forks source link

Inquiry about the signal-independent noise map calculated in the code and the setup of single image training #3

Open whtang906 opened 2 years ago

whtang906 commented 2 years ago

Hi, it's me again! I am really interested in your work! But I have found some questions about the technical detail.

According to figure 3a of the paper, you feed the signal-independent noise map with signal-dependent noise map into the network again and expect to get . I have some questions about this part.

  1. Why we need to feed the signal-dependent noise map as well? Why not just feed the signal-independent noise map and expect to get .
  2. According to the provided source code, it seems the signal-dependent noise map wasn't fed into the network? https://github.com/Reyhanehne/CVF-SID_PyTorch/blob/c89242a4ef5cc98481404192392a6f76e0fb0ec9/src/trainer/trainer.py#L87

Apart from this, can you share more detail about the setup of single image training? For example, the hyperparameters you have used (epochs, learning rate, etc) for the image shown in the appendix.

Thank you so much!

myavartanoo commented 2 years ago

Hi, thank you very much for your interest. Since the signal-dependent noise map is multiplied by the zero clean image, it dose not have any effect but still there can be an unknown signal-dependent noise map (without any effect). In other words the input to the network is 0+0^{\gamma}*N_d+N_i which is equal to N_i.

We use the same hyperparameters. We will provide the code and training details for training on the single image and other datasets very soon.

whtang906 commented 2 years ago

Ah I see. I get your point now.

Good to hear that! Thank you so much for your effort.