NVlabs
Clarification about frame size and pre-trained datasets
Hi @amritamaz and team ,
Thank you for publishing your work, reading your paper and source code has been extremely insightful!
I had a question regarding the results published in the paper, compared to what I'm seeing in the published pre-trained models: in the paper, as well as the training_metrics.json in the queen_compressed_flame_steak dataset, you're describing frame sizes in the ~0.5 MB range.
Yet, when looking at the actual files I'm seeing that most of the PKL files are in the ~26 MB range.
Could you help me understand how to interpret these sizes? In particular, do the PKL files require some additional compression to reproduce the results in the paper?
Thanks for raising this. You're right that the saved files in the current code release are significantly larger than the sizes reported in the paper.
In the paper, we compute the compressed size based on the raw bitstream lengths of the quantized tensors, without including any additional storage overhead -- you can see how this is computed in the size() function at scene/gaussian_model.py:797. The size() function replicates the computation used in the paper and gives the correct bitstream size for comparison. The current release saves these tensors using Python dictionaries for convenience, which introduces significant overhead and does not reflect the true compressed size. We're working on implementing the full functionality to save and load just the raw compressed bytestreams.
In the meantime, you can use the size() function to compute the compressed sizes directly.
