What makes this faster than instant-ngp precisely?

[kwea123]

According to the paper, I couldn't find obvious evidence that makes your method faster than instant-ngp, as you stated that most accelerating techniques are borrowed from them. Do you have any special technique that is not described in the paper?

Or is #64 related? Maybe training on the train set you get similar result with instant-ngp only after 5min.

[liruilong940607]

Hi Kui, As pointed out by #64, it was not a fair comparison. So we did a full test in #68 under the fair comparison setting, and the results are basically identical to the official implementation. We just updated the docs to reflect that.

For runtime, though we can converge to the same quality within the same amount of time, we actually trained for less iterations (20k) compared to the official impl (35k). In other words, nerfacc still runs slower at each iteration because of the python overhead (which I think is the tradeoff for flexibility), but somehow converge more efficiently.

We didn't investigate it too much as replicating NGP is not the focus here. Maybe it has something to do with the training recipes we used for NGP, which is slightly different than the paper.

On the CUDA side, there are also some minor design differences I can share with your.

• One is that in nerfacc each sample comes from the mid point of the interval (t_start, t_end), instead of just the t_start as in the NGP impl. This gives slightly better perf (like ~0.2db). The purpose of this design is to support for Mip-NeRF line of works in the future.
• Another is that in nerfacc we try to minimize the computation need to be done in CUDA kernels that are paralleled across rays (we mentioned it in the arXiv report). For example something like this is better than something like:

# this function is paralleled across rays
visibility, packed_info_visible = render_visibility(
    packed_info, t_starts, t_ends, sigma, early_stop_eps, alpha_thre
)

Nevertheless, I think there is much space for improvement in this codebase in terms of efficiency. Welcome to contribute if you are interested!

[kwea123]

thanks, so I guess in terms of instant-ngp, this is almost a re-implementation in python.