too high psnr

[hot-bulgogi]

I conducted experiments on the flame_steak sequence using 300 frames in --eval mode. Although the paper mentions achieving PSNRs of 33.54/34.30, in my case, I recorded a PSNR of 35.10 in stage 1 and 35.92 in stage 2.

I used CUDA 11.7 with an RTX 3090.

Could it be that the test set was not properly used even though I used the --eval option?

Here are a few possible reasons I am considering:

Should I initialize frame 0 using --eval even for vanilla-3DGS?
Should I remove the test views in COLMAP as well?

[SJoJoK]

Regarding your second concern, yes, we also turned on --eval when training 3DGS for frame 0. So I think that's why you get a higher PSNR for subsequent frames.

[hot-bulgogi]

Thanks a lot! I will try experiment with frame 0 using --eval.

[SJoJoK]

Glad to help:)