megengine / pmrid Goto Github PK

oppo data has 24bit for per channel:

why load raw dng with np.uint16 format?

def read_array(path: Path) -> np.ndarray:
    return np.fromfile(str(path), dtype=np.uint16)

What do anchor and V mean in class KSigma?

Hi：
Thanks for this great work. In the paper you said: Since our proposed denoising network needs to be trained for specific sensors. So, we need train sifferent model for different sensor?
But, in my opinion. Eqn. (10) indicates that the model can be used for any sensor after k-Sigma Transform

Hi, I have download the reno10x_noise dataset to estimate the k-sigma parameters, but I found that there are two folders have the same ISO=3938, is there anything wrong here?
Another question is that I found the E(x)-Var(x) curve is not always linear. When the luminance is higher, the curve is more non-linear, I want to known if I miss some pre-processing. Now I just read the raw data and minus the black level. Also I tried to use different strategy to sample the data and cannot achieve the result of yours, is there any trick here?

您的文中说：
“we choose the 10s and 30s
long-exposure subset captured by the Sony α7s II camera, and manually take
out those with visible noise, leaving 214 high quality RAW images”
“According to our noise model described in Section 3, if clean RAW images
were available, we can synthesize noisy images by sampling from a Poisson�Gaussian distribution with estimated noise parameters measured from the target
sensor.
”
现在能找到的只有曝光10s和30s的SONY数据集中的clean raw图，他们的ISO是可以知道的
但是添加噪声的输入图需要自己制作，这需要制作一个包含ISO的json文件，这有点麻烦
可以分享您的添加噪声的数据的制作代码 和 包含训练raw图参数（如ISO等）的json文件吗

Hi,
First, thank you for your work, it's very interesting!
Second, I am curious about the setup for taking noise estimation data (reno10x_noise). The paper describes a grayscale chart

But the images in reno10x_noise look like a uniform gray chart with some sort of illuminated circle

Can you please clarify?

Hi, may I know how you did the random brightness and contrast adjustment for training data? Could you recommend a method and adjustment range? Thanks.

I have a doubt regarding the UNet-5G, UNet-21G comparison with your model.(Table 1)
Could you tell me tell about the number of parameters in this model or/ and tell(share) the code of how you calculated MACs in your model?
It would be great if you could update it here.

Thanks
Nisarg

according to Eqn. (9) in the paper, variance over f(x) is x*/k+sigma/(kk) = u+sigma/(kk) .
it seems this variance dependents on k and sigma or sigma/(kk).
obviously k and sigma both are dependent on ISO.
for sigma/(kk), because k = ga, sigma = ggsigmaD+sigmaR,
so sigma/(kk) = sigmaD/(aa) + sigmaR/(aagg), which is also dependent on g, aka. ISO.

i also tried to calculate sigma/(k*k) over different iso value, using "class KSigma" in the source code with estimated parameters provided in the code: K_coeff=[0.0005995267, 0.00868861],
B_coeff=[7.11772e-7, 6.514934e-4, 0.11492713],
and find it is not Constant.

so How to understand k-Sigma transform can make signal ISO-Independent?

The encoder and decoder block are similar in paper, but i find in code, there is a relu at the end of encoder block. But decoders dont have relu at the end. why?

states = pickle.load(f) error

UnpicklingError: A load persistent id instruction was encountered,
but no persistent_load function was specified.

for x in json.load(f)
TypeError: object() takes no parameters

https://github.com/megvii-research/PMRID/blob/8ebb9e8e96559881dee957f34243933c5beb77dd/run_benchmark.py#L106

I think this is the only training script from the author of official paper.
But it didn't clarify some setting parameters, including NoiseProfile - K and B, DataAugOptions - iso_range.
And what is meaning for item.g_mean_01 from CleanRawImages?
It's really hard to reproduce the training procedure, can someone help me?

Thanks for your great work! The k-sigma parameter in the code seems to be the same for whole CFA. Is it better to use 3 different parameters for R/G/B arrays?

Could you open source the code of train script

hi ,i have a question about the input dataset range ,i have see your paper released model input range is(0,256),if i change the input range to(-1,1) or (0,1)，i am confused about the PSNR change ,can you have a try in different training data range ?

数据集下载好了，为什么我运行python3 run_benchmark.py --benchmark /path/to/PMRID/benchmark.json models/mge_pretrained.ckp没反应

一直卡着

Hello,
Thanks for your great work done. When I run the run_benchmark.py, I encountered the following error. Can you help? I am using python=3.6.

can you please provide a script for testing on our own images or let me know how to test the model on my images.

Hi，I check your code for k_sigma transform. And I find in sigma process，cvt_b compute by difference of offset and divide by k_a. I think may be we need not divide by k_a. Is there any consideration about this. Thanks.

cvt_b = (sigma / (k ** 2) - sigma_a / (k_a ** 2)) * k_a