Hi,
I have come across a few differences in the implementation proposed in the paper and the code released here.
- For one, the paper seems to mention using multi-resolution hash grids for the SDF. But a standard frequency-based positional encoding seems to be used here?
- The config files for shape optimization show a total of 300,000 steps (material optimization - 100,000), whereas the paper implementation suggests using a total of 500,000 steps.
Currently, the PSNR values I'm getting at the end of the pipeline on the NeRF Synthetic dataset are off by as much as 8+dB (in favor of the paper). Could you please share a list of the differences if one wants to reproduce the results in the paper (or get within the same ballpark)?
Thanks
Hi,
I have come across a few differences in the implementation proposed in the paper and the code released here.
Currently, the PSNR values I'm getting at the end of the pipeline on the NeRF Synthetic dataset are off by as much as 8+dB (in favor of the paper). Could you please share a list of the differences if one wants to reproduce the results in the paper (or get within the same ballpark)?
Thanks