Free Frequency Regularization (FreeNeRF) in 💯 lines of PyTorch code | NeRF#8

Machine Learning: Implementation of the paper "FreeNeRF: Improving Few-shot Neural Rendering with Free Frequency Regularization" in 100 lines of PyTorch code.
Link to the paper: arxiv.org/abs/2303.07418
GitHub: github.com/MaximeVandegar/Pap...
Udemy course: www.udemy.com/course/neural-r...
Twitter: / paperscode
-----------------------------------------------------------------------------------------------------
CONTACT: [e-mail]
#python #pytorch #nerf #neuralnetworks #machinelearning #artificialintelligence #deeplearning #data #research #neural #function #relu #positionalencoding #neuralrendering #rendering #neuralradiancefields #deeplearning #fastnerf #kilonerf #autoint #squeezenerf #3dreconstruction #novelviewsynthesis #instantngp #nvidia #radiance #fields #pleoctree #pose #estimation #learning #camera #parameters #colmap #realtime #real #time #freenerf #sparse #inputs #few #views

[@ylab3891]

thank you for your great work
When looking at freenerf-related codes on Github, it seems that background regularize is handled in one line. I don't understand this. could you please explain? Again thank you for the nice video and great work.

[@papersin100linesofcode]

Thank you for your question. I explain it here: ua-cam.com/video/-SWvlm1naQY/v-deo.html. I hope it helps :)

[@ddo_bar]

Hi, thank you for your greadt videos. I have a question. Is the .pkl data(which is included in your github) only for the "yellow loader" ? or can I use the data to train for general objects ?

[@papersin100linesofcode]

Hi, thank you for your question. I have generated it only for the yellow bulldozer, but similar pkl files could be generated for other datasets.

[@ddo_bar]

I see. thank you for answering !!

[@erfanshayegani3693]

Hi, great content! Thanks a lot for the videos.
I just have a question; which of these papers gives me a much shorter training time on a normal setup like Colab? I tried the vanialla NeRF, and it almost takes 16 hours to train for 16 epochs!!

[@papersin100linesofcode]

Hi, thank you for your question. I saw your email, I am sorry for not answering earlier. You can start with this video, or my latest one. As it is using only eight (or four) images, training is much faster. Otherwise, you may be interested in my video about K-planes which is faster both for training and inference.

[@erfanshayegani3693]

@@papersin100linesofcode thank you so much!