A novel batch normalization layer adapted to SPD networks, with a clear mathematical derivation and experimental validation. Additional results regarding convergence vs. learning rate would be interesting.