search

xmeng525 / RealTimeDenoisingNeuralBilateralGrid

[EGSR2020] Real-time Monte Carlo Denoising with the Neural Bilateral Grid
MIT License
66 stars 13 forks source link

Question about Tungsten dataset #6

Closed ChampagneBoom closed 3 years ago

ChampagneBoom commented 3 years ago

Hello there, Would you mind releasing your temporal accumulated color image of Tungsten dataset, just like the BMFR dataset https://www.dropbox.com/sh/qrf9mwofvy8ikan/AACMvXs_7W2nEk5LpiT1TSHca?dl=0? Because the hyper-parameters you used in temporal accumulation operation for Tungsten scenes is not known. It will be good if you release the accumulated dataset. Thanks!

xmeng525 commented 3 years ago

Hi, As I said in the paper, the Tungsten dataset rendered at 64-spp does not use temporal accumulation as a pre-process. I don't have the accumulated color image of Tungsten dataset. The repo I used for generating the accumulated color image is: https://github.com/maZZZu/bmfr. You can generate the accumulated dataset with the code. The camera information required for the accumulation process (the "camera position" and "lookat" vectors) could be found in https://github.com/xmeng525/TungstenFlythrough) I can help in this part. Please email me if you would like to collaborate.

ChampagneBoom commented 3 years ago

Sorry, it's my fault, I missed that part when reading the paper. Thanks for your patient explaining.

ChampagneBoom commented 3 years ago

Hi, there missed bathroom2 scene in https://www.dropbox.com/sh/54fa7cv4k3847b0/AAAEGAGxEx_TW0lGlcmmoqlua?dl=0. Would you mind updating it？

xmeng525 commented 3 years ago

Hi, I didn't generate the bathroom2 dataset.

ChampagneBoom commented 3 years ago

Hi, your paper says "Our dataset includes rendered images at a resolution of 1280×720 from nine scenes" and top left scene in Figure3 seems "bathroom" scene, and your another project https://github.com/xmeng525/TungstenFlythrough includes a "bathroom" json. maybe this misslead me. So actually you used eight scenes totally and there is a "staircase" scene instead of "bathroom" scene?

xmeng525 commented 3 years ago

Hi, We generated the json files for the bathroom scene and we did generate the rendered images using Tungsten. However, the scene is limited in area and complexity. We only generated 20 viewpoints for the bathroom and images we got from the 20 viewpoints are very similar. We decided not to include the bathroom scene in the training process. And I no longer have the images for this scene on my side.

ChampagneBoom commented 3 years ago

I am clear now, thank you very much!