Vision Mamba Discretization Study Compares Six Schemes

A recent study looked at six different methods for discretizing the Vision Mamba state space model (SSM). It found that polynomial interpolation and higher-order hold techniques really boost accuracy in tasks like image classification, semantic segmentation, and object detection, even though they take longer to train. The bilinear/Tustin transform also shows consistent improvements. The research compared zero-order hold, first-order hold, bilinear/Tustin transform, polynomial interpolation, higher-order hold, and the fourth-order Runge-Kutta method against standard visual benchmarks. Interestingly, the default zero-order hold method was noted to hurt temporal accuracy in dynamic visual situations.