Post Optimization #84
Description
I noticed that the paper mentions Post Optimization (Optional) and states, 'The learning rates are set as follows: 1.6e-4 for position, 5e-3 for scale, 1e-3 for rotation, 5e-2 for opacity, 2.5e-3 for color, and 5e-3 for camera pose.' However, in the Anysplat repository code, in src/post_opt/simple_trainer.py, I saw that the config defines: lr_means=1.6e-4, lr_scales=5e-3, lr_quats=1e-3, lr_opacities=5e-2, lr_sh=2.5e-3, and pose_opt_lr=1e-5, which does not match the 5e-3 for camera pose mentioned in the paper. Should the paper be using 5e-3 for camera pose or 1e-5, or have I misunderstood?