search

sxyu / svox2

Plenoxels: Radiance Fields without Neural Networks
BSD 2-Clause "Simplified" License
2.79k stars 360 forks source link

Is this method sensitive to training image resolution? #78

Closed Wuziyi616 closed 2 years ago

Wuziyi616 commented 2 years ago

Hi, thanks for open source such great work! I was trying the training code on some custom data, which are very simple CLEVR like scenes. However, I find the model prediction quite noisy. I notice that the lego images are 800x800 while my images are 128x128. So I try to reduce the voxel resolution from 256 to 128 and get a better PSNR, but still noisy result. Further reducing from 128 to 64 gives worse PSNR but less noisy points. Do you have any suggestions here? Sample images in my dataset are like:

rgba_00006 rgba_00010 rgba_00025

Model prediction using voxel size=64, 128, 256: 64 128 256

Wuziyi616 commented 2 years ago

Okay it seems that the main problem is the background color. The lego images have a white background with value 1, while my images have 0. Changing it to 0 solves my problem (tho still not perfect due to camera issues but much better). Increasing my image resolution to 800 further enhances the performance. 1