Training on the NeRF-DS dataset

[shuaiguouu]

I train SC-GS on the NeRF-DS dataset with this command:

python train_gui.py \
--source_path /home/gs/datasets/NeRF-DS/bell_novel_view \
--model_path outputs/bell_novel_view \
--deform_type node --node_num 512 --hyper_dim 8 \
--eval --gt_alpha_mask_as_scene_mask \
--local_frame --resolution 1 --W 800 --H 800

I can't get a good PSNR result. Is there something wrong?

[wcwac]

The author said that the old version of the code given here (https://github.com/yihua7/SC-GS/issues/18#issuecomment-2010280844) can reproduce the results of the paper with the following command , but the link is now inaccessible. --deform_type node --node_num 4096 --deform_lr_scale 1 --eval --resolution 1 --gt_alpha_mask_as_dynamic_mask --node_warm_up 500 --iterations_node_sampling 7500 --iterations_node_rendering 10000 --oneupSHdegree_step 5000 --warm_up 3000 --hyper_dim 8 --local_frame --pred_color --pred_opacity --dynamic_color_warm_up 20000 --deform_downsamp_strategy direct

[shuaiguouu]

The author said that the old version of the code given here (#18 (comment)) can reproduce the results of the paper with the following command , but the link is now inaccessible. --deform_type node --node_num 4096 --deform_lr_scale 1 --eval --resolution 1 --gt_alpha_mask_as_dynamic_mask --node_warm_up 500 --iterations_node_sampling 7500 --iterations_node_rendering 10000 --oneupSHdegree_step 5000 --warm_up 3000 --hyper_dim 8 --local_frame --pred_color --pred_opacity --dynamic_color_warm_up 20000 --deform_downsamp_strategy direct

Thank you very much! I'll try it!

[wcwac]

The author said that the old version of the code given here (#18 (comment)) can reproduce the results of the paper with the following command , but the link is now inaccessible. --deform_type node --node_num 4096 --deform_lr_scale 1 --eval --resolution 1 --gt_alpha_mask_as_dynamic_mask --node_warm_up 500 --iterations_node_sampling 7500 --iterations_node_rendering 10000 --oneupSHdegree_step 5000 --warm_up 3000 --hyper_dim 8 --local_frame --pred_color --pred_opacity --dynamic_color_warm_up 20000 --deform_downsamp_strategy direct

Thank you very much! I'll try it!

please try to replace def readNerfiesCameras(path, inter_valid=True): with def readNerfiesCameras(path, inter_valid=False):, in https://github.com/yihua7/SC-GS/blob/26cd57d09598b2f5d951029808a5ac9f0ff4f626/scene/dataset_readers.py#L517C5-L517C23 and make sure your data set file structure organized like this: dataset/NeRF/basin_novel_view/

[yihua7]

Hi, Happy to see the discussion. I have updated the sharing link of the old version code: https://pan.baidu.com/s/1Ysa3ehDmeOX5EV6insK4dg?pwd=1234. Hope the code helps as a reference.

[shuaiguouu]

Thanks for your time and patience, thank you very much!

[zhenyuan1234]

Hi, I followed the given instructions to train the HooK in the D-NeRF dataset (python train_gui.py --source_path data/d-nerf/hook --model_path outputs/hook -- deform_type node --node_num 512 --hyper_dim 8 --is_blender --eval --gt_alpha_mask_as_scene_mask --local_frame --resolution 2 --W 800 --H 800). We get a result of 39.66 PSNR, which is inconsistent with the 39.87 PSNR inside the paper, can you tell me why this could be? Thanks!

[yihua7]

Hi, If you strictly follow the readme instructions, the results won't vary too much. The two values are quite close and may differ from machines, seeds, and environments. If you are performing some comparison experiments, please feel free to fill in the results obtained on your devices.