💻 computer science

Beyond ZOH: Advanced Discretization Strategies for Vision Mamba

This paper systematically evaluates six discretization schemes within the Vision Mamba framework and demonstrates that while higher-order methods like polynomial interpolation and higher-order hold offer the highest accuracy, the bilinear (Tustin) transform provides the most favorable trade-off between precision and efficiency, making it a superior default choice over the standard zero-order hold.

Original authors: Fady Ibrahim, Guangjun Liu, Guanghui Wang

Published 2026-04-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fady Ibrahim, Guangjun Liu, Guanghui Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to "see" the world. The robot doesn't see a continuous, flowing movie like we do; instead, it sees the world as a rapid series of still photographs (frames) taken one after another.

The paper you shared is about a specific problem in how these robots process those photos. It's like trying to guess what happens between the frames.

The Old Way: The "Freeze-Frame" Robot (ZOH)

Currently, most advanced vision robots (called Vision Mamba) use a method called Zero-Order Hold (ZOH).

Think of this like a flipbook animation where, between two pages, the image just freezes.

Page 1: A ball is at the bottom of the screen.
Page 2: The ball is at the top.
The Robot's View: The robot assumes the ball stayed at the bottom for the entire time until it suddenly "teleported" to the top.

This works okay for simple things, but in the real world, things don't teleport. They move smoothly. They accelerate, they curve, and they change texture. By assuming the image stays "frozen" between snapshots, the robot gets confused about edges, textures, and motion. It's like trying to draw a smooth curve using only square blocks; it looks jagged and inaccurate.

The New Idea: Smarter Guessing

The authors of this paper asked: "What if we taught the robot to guess the movement between the photos more intelligently?"

They tested six different "guessing strategies" (mathematical methods) to see which one helps the robot understand the world better without slowing it down too much.

Here are the characters in their story:

The Freeze-Frame (ZOH): The old way. Simple, but inaccurate.
The Straight-Line Walker (FOH): This robot assumes that between two photos, the object moves in a perfectly straight line. It's better than freezing, but real life isn't always straight.
The Smooth Curve Artist (BIL - Bilinear/Tustin): This is the star of the show. Instead of freezing or drawing straight lines, this method uses a "trapezoid" trick. It looks at the previous photo and the current one and draws a smooth, gentle curve to connect them. It's like a skilled animator who knows exactly how a ball arcs through the air.
The Complex Painter (POL & HOH): These are the over-achievers. They use complex math (polynomials) to predict the movement with extreme precision. They can draw the most beautiful, accurate curves. However, they are very slow and take a long time to learn because they are trying to calculate too many details.
The Super-Computer (RK4): This is a high-precision method used for very fast, chaotic movements. It's incredibly accurate but requires so much brainpower that it often slows the whole system down too much to be practical for everyday use.

The Big Discovery

The researchers ran these robots through three major tests:

Identifying Objects: "Is that a cat or a dog?"
Drawing Boundaries: "Where exactly does the road end and the grass begin?"
Finding Things: "Where is the car in this crowd?"

The Results:

The Over-achievers (POL & HOH) got the highest scores. They were the most accurate. But, they were slow to train and required more computing power.
The Straight-Line Walker (FOH) didn't do much better than the old Freeze-Frame method.
The Smooth Curve Artist (BIL) was the perfect balance. It wasn't quite as accurate as the over-achievers, but it was significantly better than the old method, and it was almost as fast and efficient.

The Takeaway

The paper concludes that we should stop using the "Freeze-Frame" method (ZOH) for these vision robots.

Instead, we should switch to the Smooth Curve Artist (BIL). It's like upgrading from a pixelated, blocky video game to a high-definition, smooth animation. You get much clearer pictures and better understanding of the world, without needing to buy a supercomputer to run it.

In short: By teaching the robot to "fill in the gaps" between photos more smoothly, we make it see the world much more clearly, just like a human does.

1. Problem Statement

Vision Mamba (ViM) and other State Space Models (SSMs) have emerged as efficient alternatives to Vision Transformers (ViTs), offering near-linear scalability and hardware efficiency. However, current ViM implementations rely on the Zero-Order Hold (ZOH) method for converting continuous-time systems to discrete-time sequences.

The Limitation of ZOH: ZOH assumes that input signals remain constant between sampling instants. In the context of visual data, this is a significant oversimplification. Images contain continuous variations, sharp edges, textures, and structured frequency content.
Consequences: The ZOH assumption introduces discretization lag and truncation errors. Over long sequences of image patches, these errors accumulate, leading to smeared edges, loss of texture sharpness, and degraded spatial fidelity, ultimately limiting the attainable accuracy of SSM-based vision models.

2. Methodology

The authors propose a systematic, controlled comparison of six discretization schemes integrated into the Vision Mamba framework. To ensure a fair comparison, they implemented all methods within a unified, hardware-aware CUDA execution framework. This framework preserves identical memory access patterns, kernel fusion, and scan-based recurrence, isolating the impact of the numerical discretization method from architectural changes.

The six evaluated methods are:

Zero-Order Hold (ZOH): The baseline used in original ViM (assumes constant input).
First-Order Hold (FOH): Linear interpolation between samples (models smoother transitions).
Bilinear/Tustin Transform (BIL): A signal processing technique using a trapezoidal rule to map continuous analog filters to digital filters, preserving stability and high-frequency features.
Polynomial Interpolation (POL): Fits higher-order polynomial curves through sample points to capture curvature and non-linear transitions. It utilizes a non-causal approach (using future samples), aligning with ViM's bidirectional scanning.
Higher-Order Hold (HOH): Generalizes FOH using higher-degree polynomials to approximate complex temporal variations.
Runge–Kutta 4 (RK4): A fourth-order iterative numerical integration method for high-precision dynamics.

Experimental Setup:

Models: ViM (ViM-Ti) architecture.
Datasets: ImageNet-1k, CIFAR-100 (Classification); ADE20K (Semantic Segmentation); MS COCO (Object Detection).
Hardware: 4x A100 GPUs.
Constraints: No changes to model capacity, training schedules, or hyperparameters; only the discretization logic for state updates ( $\bar{A}$ and $\bar{B}$ matrices) was modified.

3. Key Contributions

Formal Characterization: Identified and formally characterized the limitations of ZOH in visual modeling, demonstrating how it degrades temporal and spatial fidelity.
Unified Benchmark: Provided the first hardware-consistent benchmark of numerical integration methods for state-space vision models, evaluating five advanced strategies against the baseline within a single, optimized CUDA kernel environment.
Design Recommendation: Established discretization as a "first-class" design decision rather than a fixed implementation detail, offering empirical justification for adopting the Bilinear Transform (BIL) as the new default baseline for SSMs.

4. Results

A. Image Classification (ImageNet-1k & CIFAR-100)

BIL: Achieved 71.95% Top-1 accuracy on ImageNet-1k (a 0.83% gain over ZOH) with similar convergence speed (232 epochs).
POL & HOH: Delivered the highest accuracy (72.76% and 72.74% respectively), representing ~1.6% gains. However, they required significantly more training epochs (272 and 289) to converge due to non-causal dependencies and higher-order coupling.
RK4 & FOH: RK4 showed modest gains but struggled with optimization stability. FOH offered negligible improvement over ZOH.
CIFAR-100: Confirmed trends, with HOH showing a ~2.32% gain over ZOH.

B. Semantic Segmentation (ADE20K)

BIL: Improved mIoU to 43.9 (a 2.9 point gain over ZOH's 41.0).
POL & HOH: Achieved the best performance with mIoU scores of 45.4 and 45.2 (gains of 4.4 and 4.2 points).
Insight: Higher-order methods better capture fine-grained spatial transitions and boundary information crucial for segmentation.

C. Object Detection (MS COCO)

BIL: Improved $AP_{box}^{75}$ from 49.6 (ZOH) to 52.1 (a 2.5 point gain).
POL & HOH: Achieved scores of 53.2 and 54.1, respectively.
Conclusion: Advanced discretization significantly enhances object localization and bounding box refinement.

D. Efficiency and Statistical Significance

Statistical Significance: Using a one-sided paired permutation test ( $\alpha=0.05$ ), BIL, POL, and HOH showed statistically significant accuracy improvements ( $p < 0.001$ ). FOH and RK4 did not meet the significance threshold.
Inference Speed:
- BIL demonstrated the best balance, offering the lowest latency at batch size 1 (13.46 ms) and highest throughput at batch size 512 (4,804 img/s).
- POL and FOH showed competitive throughput at intermediate batch sizes (4–256) but were less robust at extremes.
- Trade-off: While POL and HOH offer peak accuracy, they incur higher training time and computational overhead. BIL offers the most favorable trade-off between precision and efficiency.

5. Significance

This paper fundamentally shifts the perspective on SSM architecture design. It proves that numerical discretization is a critical hyperparameter that directly influences model performance, stability, and generalization in vision tasks.

Practical Impact: The study recommends Bilinear Transform (BIL) as the new default for Vision Mamba and similar SSMs. It provides a "sweet spot" of statistically significant accuracy gains (~0.8–2.9% across tasks) with minimal computational overhead and fast convergence.
Theoretical Impact: It validates that treating visual data as continuous signals (via higher-order or bilinear methods) rather than discrete, constant steps (ZOH) is essential for capturing the true nature of spatial frequencies, edges, and textures in deep learning models.
Future Direction: The work opens a new avenue for research into optimizing the mathematical foundations of SSMs, moving beyond the "black box" of fixed discretization rules to adaptive, task-specific numerical integration strategies.