search

yihua7 / SC-GS

[CVPR 2024] Code for SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes
https://yihua7.github.io/SC-GS-web/
MIT License
505 stars 28 forks source link

Results with ProgressiveBandFrequency #47

Closed zhenyuan1234 closed 4 months ago

zhenyuan1234 commented 4 months ago

When I try to train with the ProgressiveBandFrequency function on D-Nerf dataset, the test results are very poor, what is the reason for this? Thank you! image

yihua7 commented 4 months ago

Hi, Have you updated the frequency band of the mask at each iteration by invoking the update func? Could you check the mask value during the test time? Actually, it should perform the same as the freq embedder when all bands are on. There seem to be some bugs in your experiments. The progressive-banding option is not turned on during our experiments but left in the code, it may cause some bugs due to the lack of code maintaining.

zhenyuan1234 commented 4 months ago

Thanks for your kind reply. I made sure the mask was turned on by adding --progressive_brand_time to the training instruction. I also print the mask update process and it shows to be correct. It is worth noting that the validation metrics during training are also correct, but the metrics with python rander.py are very poor. Can you think of any reason for this, thanks!

yihua7 commented 4 months ago

The problem may be the band is not all on during running the render.py, since the behavior of train.py seems normal. Could you print the band mask during the rendering in render.py?

zhenyuan1234 commented 4 months ago

Thank you for your sincere reply. I have another question. When training on the Lego dataset using the instructions you gave, the experimental results are not good (about 25.6 PSNR). What is the reason for this, thanks!

yihua7 commented 4 months ago

Hi, please refer to the paper. Lego's train and test sets are inconsistent (the motion does not match).