easily overfitting with random init

[NeutrinoLiu]

Hi guys, when i run with DyNeRF dataset and google immersive dataset. I notice that if i set the init_type as random, the model quickly goes to overfitting after 5000 iters. while if using sfm as init_type, the model could converge to a good psnr. Can anyone explain the reason behind this?

[charisman11212]

因为sfm提供了比较好的先验，比如高斯球会初始化在物体表面，如果采用随机的话，还需要等高斯球慢慢移动或者克隆分裂过去。

Because sfm provides a better prior, for example, the Gaussian ball will be initialized on the surface of the object. If random is used, you need to wait for the Gaussian ball to slowly move or (clone / split).