why m_nerf.training.optimize_{}s variables are all fasle

[neilgogogo]

I am sorry I am a little confused with optimize_{}s variables. Why the optimize_{}s are all false, and will the training result be better if there are all true.

instant-ngp/include/neural-graphics-primitives/testbed.h

Lines 489 to 492 in 3d15eb4

	bool optimize_distortion = false;
	bool optimize_extrinsics = false;
	bool optimize_focal_length = false;
	bool optimize_exposure = false;

Kind regards

[Tom94]

Hi there,

these variables (in particular optimize_distortion and optimize_extrinsics) can help with training real-world footage that has suboptimally estimated camera parameters (such as from an online SLAM algorithm).

For all our paper results, we used either synthetic data (known camera parameters) or data with very good camera parameter fits (through offline COLMAP). Thus, fine-tuning of the camera parameters rarely gave a meaningful improvement and wasn't worth the extra compute.

That's why these are off by default -- but they can help a lot when your training data has slightly incorrect poses or lens distortion params.