This is a minimalistic implementation of paper Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains for the purpose of understanding and experimenting with fourier feature encoding in 2D which is main invention behind NeRF's success.
| Input | Learned (L=256,M=10) |
|---|---|
 |
 |
For positional encoding we can experiment with number of frequencies and multiplier to choose the best values for highest psnr. L defining number of frequencies and M being frequency multiplier. Best result was achieved using L=256 and M=10
 |
 |
|---|---|
| With L=16 & M=10 | With L=64 & M=10 |
 |
 |
|---|---|
| With L=256 & M=1 | With L=256 & M=50 |