Descriptor loss: We tested descriptor loss using different methods, including dense method (as paper but slightly different) and sparse method. We notice sparse loss can converge more efficiently with ...