Next steps

Immediate GradICON publication work

blog post on how to train

collect training scripts

Low resolution glitch- get visuals for marc

Hack on definition of loss to improve performance by reducing map compositions

move cvpr_network.py from training scripts folder into library

learn2reg

Variants of GradICON loss

shear matrix trick?

$|MM^T - id|_2^2$ as loss?

Lipshichtz constant of composition < 1 : P norm 2-> 4

Follow on papers

atlas registration

multimodal registration: synthmorph?

train one network on all the data we have- leave one out evaluation- very good learn2reg data limited

Paper: diverse regularizers

Paper: multistep SVF

Paper: inverse consistent by construction

2-D 3-D- least squares regression using 3d-3d map as "ground truth"

Neural field registration

transformer backbone? how much doed network architecture matter?

more equivariance? $\Phi(Translation \circ I^A, I^B) ?= Translation ? Phi(I^A, I^B)$