search

maincold2 / FFNeRV

FFNeRV: Flow-Guided Frame-Wise Neural Representations for Videos
MIT License
81 stars 1 forks source link

reproduction hyperparameter setting #8

Open zhaojiancheng007 opened 1 month ago

zhaojiancheng007 commented 1 month ago

thanks for your work, it's very instructive !

I am attempting to reproduce the results of FFNeRV, and I would like to ask if you have tried reducing FFNeRV's parameter count to around 3MB, similar to the settings in many other related works (HNeRV, HiNeRV, etc.). Which parameter should I adjust for this? Currently, my main approach is using a single grid with a resolution of 512 and tuning 'fc_hw_dim=9_16_39', but I’ve noticed that this configuration significantly decreases the performance.

If I want to tune FFNeRV to use fewer parameters, how should I set the parameters?

Thank you, looking forward to your reply.

maincold2 commented 1 month ago

Hi, I haven't explored the optimal settings for video representation at a lower budget, but I would like to suggest multi-resolution grids (2~4 grids, highest resolution >= 512 for 600-frame video) rather than a single grid. Actually, HNeRV and HiNeRV are later works than ours, and I noticed that HiNeRV has reported our method's performance with lower parameters, which could be helpful.

Thanks!

zhaojiancheng007 commented 1 month ago

Hi thanks for your reply.

I tried several setting, mainly around tuning fc_hw_dim and grid resolution, and achieve similiar result reported in Hinerv(bunny dataset), I will try test on other datasets and then report the result here.

Thanks!