Xin Jin

Results 28 comments of Xin Jin

您可以在finetune阶段中将图像保存下来看看吗?看看是否有这种问题。

请参考 https://github.com/Srameo/LED/issues/10#issuecomment-1805143410

1. Normally, you can use a validation set to select the results. However, since collecting a validation set can be resource-intensive, you may simply take some inputs and then use...

Thanks for your attention on our LED! However, we are sorry for that our LED was designed on camera-specific denoising, which has mostly been deployed as device-side algorthims. So LED...

Please refer to https://github.com/Srameo/LED/issues/13#issuecomment-1829882406

There are many reasons that can cause blur, but the most likely is that your GT and Input are not aligned. If you can provide an example, we should be...

The example of the input (after amplified by the ratio) and the output, or maybe both? It would be better if these RAW images are postprocessed into sRGB images.

The blur might be caused by over denoise. Is the pretrain ratio matches the testing ratio?

您好! scale是通过相机meta data中的白电平`white_level`和黑电平`black_level`,即`scale=white_level-black_level`。这一部分主要负责相机量化噪声的合成部分,在我们以及ELD的噪声模型中,量化噪声被建模为 $U(-\frac{1}{2q}, \frac{1}{2q})$,其中 $q$ 为量化步长,即 `scale`。 如果您想合成低光图像,您只需修改`ratio`为您需要的范围。由于LED主要面向的任务是极暗场景的去噪任务,因此`ratio`范围较大,在合成的图像“很黑”,可能不符合您的预期,您可以适当的降低`ratio`范围。 btw, 如果LED对您的项目有帮助的话,希望您可以帮我们点个star🌟,这将是对我们进一步科研的莫大鼓励!感谢!

1. Scale 是相机 white point - black level 的值。您可以理解成是比特深度减去黑电平。 2. 我们没有尝试过在RGB上增加噪声,但是理论上RGB上的噪声模型和RAW上的噪声模型并不一致。 - 由于RAW->RGB经过了复杂的ISP过程,因此RAW上的噪声分布在变化到RGB上时遭到了破坏,直接将噪声模型加到RGB上可能效果并不会很好。 - 但是我们并不确定在低ratio的情况下,网络学到的是GT信息还是Noise信息。如果学习到的是GT信息的话(可以参考[CVPR 2023](https://github.com/haoyuc/MaskedDenoising)),可能也会奏效;反之则不奏效。 您可以尝试之后随时找我们交流~