algohunt/vqrf

Official implementation of our CVPR 2023 paper "Compressing Volumetric Radiance Fields to 1 MB"

vqrf's Issues

When will VQ-Plenoxels and VQ-TensoRF available

Hi, @AlgoHunt

Thanks for your great work. As the current repo only contains VQ-DVGO, I'd like to know when will VQ-Plenoxels and VQ-TensoRF available or do you have any plans for this.

GPUs

Thanks for sharing this great work! What kind of 8 GPUs do you use for training the model?

About Equ.5

Why $I_{l}<\theta = 1$, rather than $I_{l} > \theta = 1$

Codebook update is different from the paper?

Nice work! But I'm confused with the code about codebook update, which is here. It looks like the new vectors in the codebook are just the mean of the grid vectors in the same cluster. It’s different from the equation (9) in the paper.

Do I miss anything?

Training time

Can you please tell me what were the training time for all your experiments.

Consider using SSIMULACRA alongside PSNR

This is to my knowledge, the best image compression metric
https://cloudinary.com/blog/detecting_the_psychovisual_impact_of_compression_related_artifacts_using_ssimulacra
https://github.com/cloudinary/ssimulacra
made by the authors of JPEG XL

A Question

Can this method or a slightly improved/shifted method of this repo be implemented into a game engine such as unreal engine or unity
So it can be used on actual game assets or is this methodology only exclusive to NeRFS , that would be super useful , also if it can be combined with Nanite that would be over powered tech?

Step by step result of VQ-tensoRF

Thanks for your great work. I'd like to know the result of step by step analysis of VQ-tenseoRF(like Table.3 in your paper)?
And do you have any plans to release VQ-TensoRF code?

Evaluation

I really enjoy this great code and paper. After training, the grids became sparse so how can you define the sampling points in each ray in inference I read sample_ray which shows that each ray has a different number of sampling points.

Thank you.

Can't find the file mean.txt

Thank you for the nice work. I try to run the trainning code, but when i call
python autotask_final.py -g "0 1" --configname nsvf_4096code --dataset nsvf

error No such file or directory: './logs/nsvf_4096code/nsvf_4096code_Bike/render_test_vq_last/mean.txt'
autotask_final.py line 169: path = f'./logs/{args.configname}/{args.configname}_{scene}/render_test_vq_last/mean.txt'
should i create the file mean.txt ? or it's a config file in the repo?