GTC 2020: Fast Denoising with Self Stabilizing Recurrent Blurs
After clicking “Watch Now” you will be prompted to login or join.
Click “Watch Now” to login or join the NVIDIA Developer Program.
Fast Denoising with Self Stabilizing Recurrent Blurs
Dmitry Zhdan, NVIDIA
In this topic NVIDIA is going to discuss latest advancements in non-DL based denoising. Basing on previous work from Metro Exodus, a new method has been introduced which is based on recurrent blur too, but it has got a lot of improvements, like: better overall performance, cleaner results, specular denoising support, fast data reconstruction, better bilateral weighting and some others. Besides the algorithm overview the topics will be covered in the talk: - mipmapping of incoming radiance - is it worth it? - proper controlling of blur radius to avoid over-blurring - accurate dis-occlusion detection - specular tracking without specular motion - fast and high amplitude noise free data reconstruction of regions with discarded history - exponential versus linear accumulation. Why linear accumulation is better? - how to compute bilateral weights in spatial passes? - input signal compression - should it be used or not? - how to fight with temporal lag