Colmap and training configuration for the MPEG dataset

[shashanknsridhara]

Dear authors,
Thanks for open sourcing this excellent work.

I have been trying the code on the MPEG dataset (basketball), but I am getting very poor results. The PSNR is around 14dB. I am running the same scripts with a small tweak in parameters such as feature stream channel P set to 8 as suggested in the paper. But my results are very bad.

I am not sure, if you have changed the colmap settings for MPEG datasets. Please include more information on how to reproduce results for MPEG dataset.

[Hliii]

Thank you for your question.

I think the first thing to be checked is whether the COLMAP results are correct. You can use meshlab to visualize the input ply file. And then there are some essential parameters to check:

• voxel size: it decides the sparsity of anchors. the larger voxel size is, the anchors will be less. We recommend 0.01 for Basketball scene.
• densify_grad_threshold: it decides the threshold to densify. If the anchor number is less than 100k for basketball, densify_grad_threshold may be too large.
• batch size: the larger batch size is, the results will be better. We recommend 4 for basketball scene.