Appendix Fig. A1.

Affine architectures. Detector outputs ReLU-activated feature maps for a single image. We compute their centers of mass (COM) and weights separately for $m$ and $f$, to fit a transform $T$ that aligns these point sets. A recurrent Encoder estimates refinements to the current transform $T_i$ from moved image $m_i=m\circ T_i$ and fixed image $f$ . Decomposer predicts a one-shot displacement field (no activation) with corresponding voxel weights (ReLU), that we decompose in a weighted least-squares (WLS) sense to estimate affine transform $T$. Parentheses specify filter numbers. We LeakyReLU-activate the output of unnamed convolutional blocks (param. $\alpha =0.2$ ). Stacked convolutional blocks of decreasing size indicate subsampling by a factor of 2 via max pooling following each activation.