This is a fork of TensorRF, reimplemented in Pytorch Lightning. With this implementation, we achieve beter modulization and faster training.
Create a conda environment using environment.yml. It's tested on cuda 11.7 and corresponding torch version.
To train
python launch.py --config ./configs/lego.txt --train
To run testing
python launch.py --config /path/to/saved/config.txt --test --resume /path/to/ckpt
The original structure is decomposed to:
- Density field
- Radiance field
- Occupancy grid
- A renderer that implements the rendering algorithm
This decomposition makes modules reusable and could be further used for other applications such as inverse rendering, semantic segmentation.
| G.T. | Original | This repo |
|---|---|---|
 |
 |
 |
For lego scene, on a single RTX 2080, with batch size 4096 and 30000 iters
| time | psnr | |
|---|---|---|
| Original | 25:30 | 35.51 |
| PL | 22:14 | 35.53 |
Other metrics(LPIPS, SSIM) are similar to values reported in original paper.