Attempt to Reproduce 3DGS Vanilla Results Using LLFF Dataset

[hyeon9698]

First, I want to say that I found the work presented in your paper to be very impressive. I’ve been testing your approach and am working on reproducing the vanilla results as described.

To do this, I used the following command:

bash scripts/run_llff_vanilla.sh 0 /nerf_llff_data/leaves /output/leaves_3views

Below is the code snippet used in run_llff_vanilla.sh:

export CUDA_VISIBLE_DEVICES=$1
dataset=$2 
workspace=$3

python3 train.py \
--source_path $dataset -m $workspace \
--eval  -r 8 --n_views 3 \
--random_background \
--iterations 10000 --position_lr_max_steps 10000 \
--densify_until_iter 10000 \
--densify_grad_threshold 0.0005 \
--onlyrgb
# --gaussiansN 2 \
# --coprune --coprune_threshold 5 \
# --coreg --sample_pseudo_interval 1 --start_sample_pseudo 500

Could you please take a look and let me know if these settings are consistent with those used to reproduce the vanilla results in your paper? I want to ensure that I’m following the correct steps.

Additionally, I’d be happy to share the results I obtained using these settings if that would be helpful. Below, you’ll find a comparison: the image on the left shows my reproduced result, and the image on the right is from your paper.

Thank you for your time and assistance. Looking forward to your feedback.

[jiaw-z]

Hello, the code you are running is the same as mine except for the random_background setting. Regarding random_background, I discovered during my experiments that regions similar in color to the background can cause overfitting due to the presence of the background color, especially under sparse views. However, the original 3DGS, when using random_background, might reconstruct a large number of Gaussian points. Since there isn’t an operation similar to co-prune, it is not suitable to use random_background. The visualization results you reproduced look reasonable. And there might be some differences because 3DGS has significant randomness under sparse views, as discussed in the paper.

[hyeon9698]

Thank you in advance for your kind response :)

I have a question regarding the 3DGS results. The performance metrics for 3DGS in your paper are noticeably higher compared to other papers (for example in LLFF dataset; Other paper is about 17 PSNR but CoR-GS is 19.22), and I’m not sure why. When I run experiments using CoR-GS, I get similar results to yours, but I can’t figure out why they outperform other implementations in 3DGS setting.

If possible, I would greatly appreciate it if you could kindly share any insights you may have on this. Thank you so much!

[jiaw-z]

Hello, I referenced the parameters of FSGS for the vanilla 3DGS in this case. When I was reproducing the results, I obtained a score of 19.22. I’m not quite sure why my reproduced results are better than other reported results. And I tried multiple times, the results consistently stayed within a similar range around 19.22. Therefore, I reported the 3DGS results that I reproduced. However, I also found that the choice of hyperparameters in 3DGS has a significant impact on performance. When I used the default parameters from the original 3DGS for the sparse-view llff dataset, I noticed that the performance was slightly worse compared to the set of parameters from FSGS.

[hyeon9698]

Thank you for your kind response and helpful insights. I'll keep experimenting with the hyperparameters. I really appreciate your time and assistance.

Best regards, Donghyeon