Performance evaluation details

[MarekKowalski]

Hello!

The paper mentions that you get 0.015s per frame for the 256x256 model on RTX4090. Can you share more about how that is calculated?

On my end, on RTX3090, calling EncoderNoPoSplat.forward for 256x256 me 0.25s per frame, which seems to be more than the difference between the 3090 and 4090 would suggest. I run every frame separately (batch size of 1), the time is averaged over about 100 forward passes. Any thoughts on what may be leading to such a different result for me?

Thanks,

Marek

[botaoye]

Hi, thank you for your interest in our work, our reported runtime is calculated here and the results will be printed out/saved at the end of the evaluation.

Could you please share your results using this function?

[MarekKowalski]

Thanks a lot for sharing. Once I get the training code running I'll let you know what numbers I am getting there.