One of the authors here. One thing about quaternion convolution is that you can write a color image into quaternion space by considering each channel as an imag

One of the authors here. One thing about quaternion convolution is that you can write a color image into quaternion space by considering each channel as an imaginary axis. This lets the convolution act on the entire color space in a different way compared to real-valued networks, which may it do better for things like segmentation where you need to be more sensitive to changes in the color space.