- mask sky
- use the perfectly cropped point clouds for training
- do not optimize gaussian means
- not allow model to create or remove Gaussians
- enable batched optimization
- fix some cameras on the back are not optimized problem
- is everything wrongly places? because train cameras dont exactly fit?
- track the loss flow and see what happens to it
- add 5 degrees and 50cm error on all 3 axes
- use cm and degrees in error tracking file
- add NN for gaussians
- skip 1 out of 2 frames? ( results in 40 frames for each sensor)
- add errors to the left camera as well?
- check if more data on point clouds exist in kitti dataset
- fish eye camera?