Details for Comparisons #35
Description
Hello, congratulations to this great work.
Could you please provide more details about how the metrics Tbl. 1 are computed?
In particular, could you provide the scripts that were used to compute these metrics or share details on how to conduct a fair comparison, i.e. how was the output post-processed (masking, etc.), what backbone was used for LPIPS and do you use the same implementations for PSNR and SSIM as in the original 3DGS repository? For how many iterations did you train your model?
Moreover, how was the rendering speed of 300FPS measured? Is this the time it takes for a single input to pass through both the deformation model and the rasterization module or was it done for batched inputs and with asynchronous kernel dispatching (i.e. without calling torch.cuda.synchronize)
Thank you.