search

zju3dv / PGSR

[TVCG2024] PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction
https://zju3dv.github.io/pgsr/
Other
507 stars 33 forks source link

How to obtain mesh and mesh normal images from the paper #45

Open angieAAAAA opened 1 month ago

angieAAAAA commented 1 month ago

Thank you for your excellent work and for providing the code. I have a question regarding the results shown in the paper, specifically the mesh and mesh normal images.

Could you please provide more details on how to obtain the mesh and mesh normal visualizations (like the ones shown in readme)? Are there any specific steps or tools in the code that need to be used to generate these visualizations?

I would really appreciate any guidance or scripts you could share to help me reproduce those results.

danpeng2 commented 1 month ago

Hello, we use the Open3D tool to display the mesh and normals. After opening the mesh in Open3D, press the shortcut ctrl+9 to toggle the normal map, and press 9 to save the image.

angieAAAAA commented 1 month ago

Thanks for your reply. However，when running experiments on the Courthouse dataset, the generated normal maps are noticeably less smooth compared to the results shown in the paper. Could this be related to dataset processing, parameter settings, or model training issues? I would appreciate any insights or suggestions to address this. 屏幕截图 2024-10-14 112000

danpeng2 commented 1 month ago

Hello, the result is quite poor. Are your data preprocessing and training parameters consistent with the README?

angieAAAAA commented 1 month ago

I followed the instructions in the README and downloaded the Courthouse dataset from the official website, but I encountered a "CUDA out of memory" error. To work around this, I extracted 1/5 of the images from a mid-range distance. Then, I used the convert.py script from the 3DGS project to generate the corresponding camera parameters.

After that, I followed the Custom Dataset section in the PGSR README and ran the following commands:

python train.py -s data/courthouse -m output/courthouse --max_abs_split_points 0 --opacity_cull_threshold 0.05
python render.py -m output/courthouse --max_depth 10.0 --voxel_size 0.01
danpeng2 commented 1 month ago

Extracting 1/5 of the images results in overly sparse data, making high-quality reconstruction impossible. You can refer to the script scripts/run_tnt.py, and add the argument ‘--data_device cpu’ to cache the data on the CPU, preventing GPU memory overflow.

angieAAAAA commented 1 month ago

Thanks for your suggestions! I will try them out and see how it goes.