Hypernerf Practice
This README contains overview on applyiing several custom scenes to reproduce the hypernerf project.
- hand
- chess
- laptop
- dvd
- tomato-mark
hand
most successful one, trained intensively, I cropped out wobbly artifacts from top
PSNR:
coarse metrics: metric/psnr≈31.7468
fine metrics: metric/psnr≈32.5462
chess
minimal movement, trained intensively, lighting was varying a lot during capturing, causing flickering
PSNR:
coarse metrics: metric/psnr≈28.4936
fine metrics: metric/psnr≈30.5115
laptop
trained intensively, the thin screen edge failed, far field not so great. more wobbly artifacts, I assume caused by either camera registration error or hypernerf's natural artifact.
PSNR:
coarse metrics: metric/psnr≈23.8846
fine metrics: metric/psnr≈25.1817
dvd
the floating dvd logo: not so great, raw capture is not well set. too short, and "time" has been "associated" with viewing angle by the model.
PSNR:
coarse metrics: metric/psnr≈27.1150
fine metrics: metric/psnr≈28.6814
tomato-mark
wobbly artifacts, I assume caused by either camera registration error or hypernerf's natural artifact. This is present in the hand scene and laptop scene as well
interpolation:
PSNR:
coarse metrics: metric/psnr≈27.5819
fine metrics: metric/psnr≈27.9282
The Wobbly artifact mentioned earlier:
reproduction steps
Due to deprecation, the hypernerf project doesn't work out of the box.
I had to modify dependencies, then minor changes at rendering part.
For this reproduction, I only used the notebooks.
The provided instructions below have been tested to work.
Click to reproduction steps
Dataset&checkpoint Download:
All in one:
https://huggingface.co/datasets/xieyizheng/hypernerf_custom_scenes/resolve/main/custom_scenes_all.zip
Separately:
https://huggingface.co/datasets/xieyizheng/hypernerf_custom_scenes/tree/main