2 ms·
Positional Encoding
- PaulHoule 4y agoIt interests me that very similar math is used in NeRF models. It seems if you use the usual representation and the usual loss functions, the model training process thinks a blurry image is good enough so you get blurry images, that kind of positional encoding in NeRF forces the model to spend more effort on high-frequency components than it otherwise would.