I can't really comment on the novelty of this work, but I don't think the connectivity structure makes much sense.

I mean, it does in the sense that local pixels are strongly correlated and a binary tree will captures this. In fact if you add weight-sharing to the K-tree model you can recover 1D convolution with a stride and kernel of 2.

But is this really the right operation for images? Why fixed kernel of 2? I think capsules or some other vector-based operation would make more sense. Perhaps with a learned or dynamic connectivity pattern.