Partial Ring Scan: Revisiting Scan Order in Vision State Space Models

This paper introduces PRISMamba, a rotation-robust Vision State Space Model that improves accuracy, efficiency, and geometric stability by replacing fixed linear scan orders with a concentric ring-based traversal and partial channel filtering.

The Gist

Imagine you are trying to describe a complex picture, like a photo of a dog chasing a rabbit, to a friend over a phone line. You can only send one word at a time. The order in which you describe the picture matters immensely.

If you describe the picture row by row (left to right, top to bottom), you might say, "Dog's ear... dog's nose... rabbit's ear... dog's tail." If you rotate the picture 90 degrees, your row-by-row description suddenly becomes a mess. The "dog's ear" is now far away from the "dog's nose" in your list, and the rabbit's ear might end up sandwiched between the dog's tail and the background. Your friend (the computer model) gets confused because the things that belong together in the picture are now far apart in your story.

This paper, titled "Partial Ring Scan: Revisiting Scan Order in Vision State Space Models," argues that the way we "read" images in modern AI is too rigid. The authors propose a new, smarter way to read images that stays calm even when the picture spins around.

Here is the breakdown of their ideas using simple analogies:

1. The Problem: The "Snake" vs. The "Target"

Current AI models (like VMamba) often read images like a snake slithering across a grid. They go left-to-right, then drop down, then go right-to-left.

• The Issue: If you rotate the image, the "snake" path gets cut off. The AI has to jump across empty space (padding) or jump to the wrong part of the picture. It's like trying to read a book where someone rotated the page; the sentences no longer make sense because the words are in the wrong order.
• The Result: When the image rotates, the AI's performance drops because it loses track of how objects connect.

2. The Solution: The "Onion" Strategy (Ring Scan)

The authors, Yi-Kuan Hsieh and colleagues, suggest a different way to read the picture: The Ring Scan.

Imagine the image is an onion or a target board.

• Step 1: Instead of reading row by row, the AI peels the image in concentric rings, starting from the very center and moving outward.
• Step 2: Inside each ring, the AI doesn't care about the specific order (clockwise or counter-clockwise). It just gathers all the information from that ring together.
• Step 3: It then moves from the inner ring to the next ring, and so on.

Why this is a game-changer:
If you rotate the picture, the "onion" doesn't change. The center is still the center, and the outer ring is still the outer ring. The AI doesn't have to relearn the order; it just sees the same rings, just rotated slightly. This makes the AI incredibly robust against rotation.

3. The Efficiency Hack: The "VIP Lane" (Partial Channel Filtering)

Processing every single piece of information in the image is expensive and slow. The authors noticed that not all "channels" (think of these as different colored filters or types of data) are equally important. Some are noisy or redundant.

• The Old Way: The AI tries to process everything through the complex "Ring" path.
• The New Way (Partial Channel Filtering): The AI acts like a bouncer at a club. It quickly checks which channels are "VIPs" (the most informative ones).
  • VIPs get sent through the main, high-speed "Ring" path.
  • Regulars (the less important data) are sent down a simple, lightweight "backdoor" path (a residual branch).

The Analogy: Imagine a busy highway. Instead of forcing every car to take the long, winding scenic route, the smart system lets only the fast sports cars take the scenic route while the slow trucks take a direct, simple bypass. The result? The highway moves faster, and the important cars still get the detailed view they need.

4. The Results: Faster, Smarter, and Stronger

The team tested their new model, called PRISMamba, against the current leaders (like VMamba).

• Accuracy: On a standard test (ImageNet), PRISMamba got 84.5% accuracy, beating the previous best (VMamba) which got 82.6%.
• Speed: It processed images much faster (3,054 images per second vs. 1,686 for VMamba).
• Efficiency: It used less computing power (3.9G FLOPs vs. 5.6G).
• Rotation: When they rotated the images by 60 degrees, the old models dropped in performance by about 2%. PRISMamba barely noticed; its score stayed almost exactly the same.

Summary

The paper claims that by changing how the AI reads the image (from a snake-like path to a ring-based path) and by filtering out the boring data (Partial Channel Filtering), they created a model that is:

1. More accurate (it sees better).
2. Faster (it thinks quicker).
3. More robust (it doesn't get confused when the picture spins).

They call this a "low-cost, principled approach," meaning they didn't need to build a massive, expensive new brain; they just rearranged the furniture and opened a VIP lane.

Technical Summary

Technical Summary: Partial Ring Scan: Revisiting Scan Order in Vision State Space Models

Problem Statement
State Space Models (SSMs), particularly Vision Mamba variants, have emerged as efficient alternatives to attention mechanisms for vision tasks, offering linear-time sequence processing. However, these models require serializing 2D images into 1D token sequences based on a predefined scan order (e.g., raster, serpentine). The authors argue that this scan order is a critical, often overlooked factor that dictates spatial adjacency within the sequence.

Conventional fixed-path scans (raster, bidirectional, diagonal) preserve sequence-space alignment only under specific transformations like flips. Under in-plane rotations, these paths suffer from global reindexing and padding, which fracture object continuity and misalign the recurrent kernel's neighborhood. This forces the SSM to expend capacity repairing local coherence rather than modeling semantics, leading to performance degradation (1–2% Top-1 drop) and reduced robustness to geometric transformations.

Methodology: PRISMamba
The paper proposes Partial Ring Scan Mamba (PRISMamba), a rotation-robust architecture designed to align sequence adjacency with geometric structure without requiring rotation-specific training. The method consists of three core components:

1. Ring-by-Ring Alternating Scan:

   • Partitioning: The image is partitioned into concentric rings based on the Euclidean distance of pixels from the image center.
   • Intra-Ring Aggregation: Within each ring, pixels are traversed in a closed loop (alternating clockwise and counter-clockwise for adjacent rings). A lightweight selective SSM processes this loop, and the output is aggregated into a ring descriptor via averaging. This aggregation is order-agnostic; a rotation merely induces a cyclic shift in the loop, which does not alter the averaged descriptor.
   • Inter-Radial Propagation: The ring descriptors are processed from the innermost to the outermost ring using a second, short selective SSM to propagate context radially. This two-level design maintains linear-time complexity ($O(HW)$) while avoiding the fragility of global path-based serialization.

2. Partial Channel Filtering (PCF):

   • To improve efficiency, the authors introduce a hard routing mechanism that filters feature channels before they enter the recurrent ring pathway.
   • Channel salience is determined via Global Average Pooling (GAP). Channels with salience above the mean threshold are routed through the ring-wise SSM, while the remaining channels bypass this heavy computation via a lightweight residual branch.
   • This approach reduces FLOPs and activation memory while maintaining accuracy, contrasting with soft attention mechanisms (e.g., SE, CBAM) that process all channels downstream.

3. Write-Back and Fusion:

   • The radial outputs are broadcast back to the pixel grid and fused with the backbone stream via a 1×1 projection and residual connection, preserving spatial dimensions.

Key Contributions

• Systematic Scan-Order Analysis: The paper provides the first systematic study demonstrating that scan order is a "first-class" hyperparameter. It shows that common raster and serpentine scans fracture object continuity and are highly sensitive to rotation, whereas ring-based scans maintain stability.
• Rotation-Stable Traversal: The introduction of a ring-based scan scheme that groups pixels into concentric rings and performs order-agnostic aggregation. This preserves the linear-time core of SSMs while eliminating the need for polar remapping or rotation-specific data augmentation.
• Partial Channel Filtering (PCF): A novel efficiency module that routes only the most informative channels through the recurrent pathway. This reduces computational cost and improves throughput while slightly enhancing accuracy by focusing the recurrent updates on high-signal features.
• Unified Architecture (PRISMamba): An integrated model that combines these elements to achieve state-of-the-art performance among Vision SSMs.

Experimental Results
The authors evaluate PRISMamba on ImageNet-1K, MS COCO, and under rotation/occlusion stress tests:

• ImageNet-1K Classification: PRISMamba achieves 84.5% Top-1 accuracy with 3.9G FLOPs and 3,054 img/s throughput on an A100 GPU. This outperforms VMamba (82.6%, 5.6G FLOPs, 1,686 img/s) by 1.9% in accuracy while using 30% fewer FLOPs and achieving ~1.8× higher throughput.
• MS COCO Detection/Segmentation: Using a 1× schedule (1280×800), PRISMamba achieves 48.9 APbox and 43.2 APmask with 235G FLOPs, outperforming VMamba and GroupMamba while using 10–19% fewer FLOPs.
• Rotation Robustness: Under 30° and 60° rotations, PRISMamba maintains near-perfect stability (dropping only 0.1–0.2%), whereas fixed-path baselines (VMamba, GroupMamba, etc.) degrade by 1.0–2.0%.
• Occlusion Robustness: In random-mask occlusion tests, PRISMamba shows significantly lower performance drops (≤0.6%) compared to fixed-path variants (up to 1.3%), attributed to the dilution of local damage via ring averaging and radial propagation.
• Ablation: The PCF module alone reduces FLOPs by ~15% and increases throughput by ~40% while improving accuracy by 0.4%, demonstrating that hard channel filtering is more effective than soft attention for recurrent SSMs.

Significance and Claims
The paper claims that scan-order design, coupled with targeted channel allocation, is a crucial but underexplored factor for accuracy, efficiency, and robustness in Vision SSMs. PRISMamba demonstrates that it is possible to resolve the sequence–geometry mismatch in path-based models without increasing model size or requiring specialized training. The authors position their work as a principled, low-cost approach to making Vision SSMs more robust to geometric transformations and more efficient through selective channel processing.

Limitations and Future Work
The authors acknowledge that their current implementation relies on a fixed image center and discrete ring width, which may be suboptimal for off-center subjects or extreme aspect ratios. Future work includes learning ring origins and widths, designing content-adaptive partitions, and extending the ring traversal concept to spatio-temporal domains (video) and 3D volumes.