search

chenhsuanlin / bundle-adjusting-NeRF

BARF: Bundle-Adjusting Neural Radiance Fields 🤮 (ICCV 2021 oral)
MIT License
793 stars 114 forks source link

Why do you use Xavier init instead of Kaiming init for the parameters of NeRF? #21

Closed ootts closed 2 years ago

ootts commented 2 years ago

I tried to use the built-in initialization (kaiming init) and the network cannot converge. Why is the xavier init necessary?

chenhsuanlin commented 2 years ago

Hi @f-sky, we tried to follow the original TensorFlow implementation as closely as possible, which uses the Xavier initialization by default, and empirically we found it to work slightly better than Kaiming init (PyTorch default). I don't have a good answer to why one's better than the other; I'm guessing neural representations could be sensitive to initialization strategies. That said, it's weird that you didn't see convergence. Were you running BARF or the original NeRF?

chenhsuanlin commented 2 years ago

Closing due to inactivity, please feel free to reopen if necessary!