search

sxyu / svox2

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

Fail to reproduce the results (bad visualization + low PSNR) (Fixed) #35

Closed KelestZ closed 2 years ago

KelestZ commented 2 years ago

Thanks for the amazing work. When I run the code on nerf_synthetic data (i.e., Lego), I couldn't get the same results as good as shown in the paper. I am showing a rendered LEGO result as below (right): 0000 The PSNR I got is ~21 on lego. I believe I followed the instructions for the installation and running. The example command line is like:

python opt.py ../data/lego -t ckpt/lego -c configs/syn.json

I'm also repeating the experiments on other data. The Drum and the hotdog in Syn data and the fern in LLFF data seem fine. No idea about other data just yet. Some good examples I got: image image image

pwais commented 2 years ago

huh when i train lego from scratch i get good results: individualImage

Screen Shot 2022-01-03 at 4 24 31 PM

you're using a pre-trained checkpoint? maybe the code and checkpoint diverged

KelestZ commented 2 years ago

you're using a pre-trained checkpoint? maybe the code and checkpoint diverged

No, I am training it from scratch. This seems weird when I got ok results on the other data. On lego, my PSNR curve converges much slower than yours: image

sarafridov commented 2 years ago

Are you sure you're using the correct Lego data? The truck seems to be in a different position than it is in the standard NeRF-synthetic dataset; perhaps somehow multiple Lego training sets got intermixed?

KelestZ commented 2 years ago

Are you sure you're using the correct Lego data? The truck seems to be in a different position than it is in the standard NeRF-synthetic dataset; perhaps somehow multiple Lego training sets got intermixed?

Thanks for the reminder. Fixed the issue and closed the question :)