Regarding the discrepancy in experimental results of cut_roasted_beef in N3DV dataset.

[LabubuJ]

Thank you for your excellent work. I encountered some issues during my experiments.
The metric of cut_roasted_beef I got is {
"SSIM": 0.9544036984443665,
"SKSSIM": 0.9497656226158142,
"SKSSIM2": 0.9747065305709839,
"PSNR": 31.879066467285156,
"LPIPS": 0.047466572374105453,
"LPIPSVGG": 0.13830572366714478,
"times": 0.01961996965110302
}
There is a significant discrepancy from the PSNR of 33.73 provided in the paper.
I used the code and commands you provided, and I would like to know how to obtain normal results.

[LabubuJ]

The training configuration is as follows.
{
"name": "Python train cut_roasted_beef",
"type": "debugpy",
"request": "launch",
"program": "./train.py",
"console": "integratedTerminal",
"args": [
"-s", "/home/disk1/jh/E-D3DGS/dataset/cut_roasted_beef_ed3dgs",
"--config", "/home/disk1/jh/E-D3DGS/arguments/dynerf/cut_roasted_beef.py",
"--model_path", "/home/disk1/jh/E-D3DGS/output/cut_roasted_beef",
"--expname", "dynerf/cut_roasted_beef",
"-r", "2",
],
"python": "/home/disk1/jh/conda_envs/ed3dgs/bin/python"
}

[juno181]

Thanks for your experiment. This appears to be the same issue as #3. I am suspecting that the problem is that COLMAP is not producing the same results. I am currently revising the proprocessing and will update the code.

[LabubuJ]

Thanks for your response！ Looking forward to your update.发自我的手机-------- 原始邮件 --------发件人： Juno ***@***.***>日期： 2024年11月24日周日 晚上9:33收件人： juno181/Ex4DGS ***@***.***>抄送： JiaoChenGuang-JiaoHan ***@***.***>, Author ***@***.***>主 题： Re: [juno181/Ex4DGS] Regarding the discrepancy in experimental results of cut_roasted_beef in N3DV dataset. (Issue #10) Thanks for your experiment. This appears to be the same issue as #3. I am suspecting that the problem is that COLMAP is not producing the same results. I am currently revising the proprocessing and will update the code. —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: ***@***.***>

[YJ-LIAO]

Thanks for your experiment. This appears to be the same issue as #3. I am suspecting that the problem is that COLMAP is not producing the same results. I am currently revising the proprocessing and will update the code.

Did you completely revise the processing code? I saw you submit a new pre_n3d_colmap.py, but I also have a bad result in coffee_martini. The metric of coffee_martini I got is PSNR: 28.13.

[juno181]

The new version of pre_n3d_colmap.py was submitted primarily to address issues related to resolution. However, the parts related to COLMAP densification are still being worked on and haven't been completely revised yet.

[YJ-LIAO]

The new version of pre_n3d_colmap.py was submitted primarily to address issues related to resolution. However, the parts related to COLMAP densification are still being worked on and haven't been completely revised yet.

I found that the reconstruction evaluation results for dense inputs are much closer to those for sparse inputs in the paper, is this normal?

[juno181]

That's correct. This is because a dense input reduces the randomness of densification, which makes the result more stable. I am improving this by using COLMAP densification.

[TCQian]

That's correct. This is because a dense input reduces the randomness of densification, which makes the result more stable. I am improving this by using COLMAP densification.

I am facing the same issue here. Is the COLMAP densification ready yet?