1. Tested on customized blender data
2. Validate the Paper's effectiveness by comparing it with a Fully-Convolutioal-Network

1. scale = 10.0 achieves the best result
2. Fourier-Feature-Transform make features shift-invariant and tunable
3. run_mlp_batch.py mlp_batch.py Train with multiple objects from same category, does not work, try to figure out a solution

https://github.com/yenchenlin/nerf-pytorch https://people.eecs.berkeley.edu/~bmild/fourfeat/index.html