Beyond ZOH: Advanced Discretization Strategies for Vision Mamba

arXiv cs.CV / 4/23/2026

📰 NewsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper argues that Vision Mamba’s default zero-order hold (ZOH) discretization weakens temporal fidelity in dynamic visual settings and limits achievable accuracy.
It systematically compares six discretization methods within the Vision Mamba framework, including ZOH, FOH, BIL/Tustin, polynomial interpolation (POL), higher-order hold (HOH), and RK4.
Experiments on standard vision benchmarks across image classification, semantic segmentation, and object detection show POL and HOH deliver the biggest accuracy improvements but require more training-time computation.
Bilinear/Tustin (BIL) consistently improves over ZOH with relatively small overhead, making it the best precision–efficiency trade-off and a recommended default baseline for SSM-based vision models.
Overall, the work highlights discretization as a key design lever for state space model (SSM) vision architectures and provides empirical justification for method selection.

Abstract

Vision Mamba, as a state space model (SSM), employs a zero-order hold (ZOH) discretization, which assumes that input signals remain constant between sampling instants. This assumption degrades temporal fidelity in dynamic visual environments and constrains the attainable accuracy of modern SSM-based vision models. In this paper, we present a systematic and controlled comparison of six discretization schemes instantiated within the Vision Mamba framework: ZOH, first-order hold (FOH), bilinear/Tustin transform (BIL), polynomial interpolation (POL), higher-order hold (HOH), and the fourth-order Runge-Kutta method (RK4). We evaluate each method on standard visual benchmarks to quantify its influence in image classification, semantic segmentation, and object detection. Our results demonstrate that POL and HOH yield the largest gains in accuracy at the cost of higher training-time computation. In contrast, the BIL provides consistent improvements over ZOH with modest additional overhead, offering the most favorable trade-off between precision and efficiency. These findings elucidate the pivotal role of discretization in SSM-based vision architectures and furnish empirically grounded justification for adopting BIL as the default discretization baseline for state-of-the-art SSM models.