Applying CamP to mipnerf360 datasets degrades PNSR and SSIM values

[ppponpon]

Hi, first of all thanks for this great paper and code. I finally managed to get your code to work!!
First, I tried to a task of replicating Table 1 in your paper for Single-Scale mipnerf360 dataset(7samples (bicycle,bonsai,counter,garden,kitchen,room,stump)), comparing with to without CamP.
Unfortunately, the average value of psnr:30.68,ssim:0.89(without CamP) degraded to psnr:27.43,ssim:0.82(with Camp)!
I used zipnerf/360_train.sh, zipnerf/360_eval.sh and zipnerf/360_render.sh, only rewriting DATA_DIR and CHECKPOINT_DIR, for w/o CamP.
And I used camp/360_train.sh, zipnerf/360_eval.sh and zipnerf/360_render.sh, rewriting DATA_DIR and CHECKPOINT_DIR and add option "-gin_configs=configs/camp/camera_optim.gin" for zipnerf/360_eval.sh and zipnerf/360_render.sh , for with CamP.

Unit test is all OK(with JAX), and Single-Scale mip-NeRF 360 Dataset Results of zipnerf(psnr,ssim) are roughly the same or better.
What do you think could be the reason it isn't working?

[ppponpon]

i.e. At garden sample,psnr:28.28,ssim:0.8664(without CamP) ,psnr:25.92,ssim:0.7548(with CamP).
Does CamP make the alignment worse?
What values ​​did everyone get?
Am I misunderstanding the results?

[ppponpon]

The following table shows the results of the test for mipnerf360 dataset.

PSNR	bicycle	bonsai	counter	garden	kitchen	room	stump	average
w/o Camp	25.888	35.817	29.8137	28.2849	33.55909	33.91839	27.47	30.67849
with Camp	24.707	29.037	27.4499	25.9273	26.86339	31.93044	26.07	27.42672
with Camp(2nd)	24.863	29.074	27.5143	26.1528	27.05436	32.08654	26.00	27.53475

The ssim values ​​show the same trend.
Why is Camp not effective? Do you have any idea?