Add average_init_density to improve robustness of nerfacto training

[jb-ye]

This has been my personal recipe for training nerfacto: the basic issue of current nerfacto implementation is the scene is always initialized with an average density at 1.0 (due to the exponential activation of MLP). This is not a robust choice in general especially when scene depth is much higher than 1.0.

One famous example is the mipnerf360 stump dataset, when training with the recommended setting before this PR

ns-train nerfacto --data stump \
  --pipeline.model.appearance_embed_dim 1 \
  --pipeline.model.camera_optimizer.mode off

The PSNR on eval set is only 18. With this PR, one can choose to train with a multiplier on top of density output. My personal experience is to set it 0.01 so the scene is very transparent initially.

ns-train nerfacto --data stump \
  --pipeline.model.appearance_embed_dim 1 \
  --pipeline.model.camera_optimizer.mode off \
  --pipeline.model.average_init_density 0.01

The PSNR on eval set will be improved from 18 to 25.3. Similar improvement can also be observed in some of nerfstudio dataset.

At last, this feature would not produce more quality in the final render, but will avoid some bad optimization outcomes due to improper initialization.

[kerrj]

sgtm!