Research questionHow can depth compression merge padded convolutions without enlarging kernels for faster embedded inference?Padding complicates analytical layer merging, while enlarged kernels can offset the speed benefits of reducing network depth. The challenge is to merge layers efficiently without increasing kernel size.