SpaRTAN: Spatial Reinforcement Token-based Aggregation Network for Visual Recognition

Quan Bi Pay, Vishnu Monn Baskaran, Junn Yong Loo, KokSheik Wong, Simon See

TL;DR

SpaRTAN is a new lightweight CNN architecture that achieves high-performance visual recognition by focusing on "middle-order" spatial features and a novel wave-based channel aggregation method. It delivers competitive results on ImageNet and COCO benchmarks with a fraction of the parameters used by traditional models—achieving 77.7% ImageNet accuracy with just 3.8M parameters.

What Problem Does This Paper Address?

Despite the rise of Vision Transformers (ViTs), Convolutional Neural Networks (CNNs) remain popular due to their efficiency and "translation equivariance." However, modern architectures suffer from two main issues:

  1. Simplicity Bias: Models often focus on very simple, low-order features (like edges) or very global, high-order features (like background context), while neglecting the complex "middle-order" textures and shapes most discriminative for identifying objects.
  2. Information Redundancy: Many modern CNNs adopt MLP-like blocks from Transformers that require high "expansion ratios" (4x or 8x), leading to significant parameter bloat.

SpaRTAN (Spatial Reinforcement Token-based Aggregation Network) bridges these gaps by capturing multi-order spatial interactions more effectively and reducing channel-wise redundancy.

Methods

SpaRTAN utilizes a hierarchical, pyramidal structure across four stages, introducing two innovative modules:

Overview of the SpaRTAN hierarchical architecture across four stages.
  • Spatial SMixer: Instead of using a single large convolutional kernel, SMixer uses a two-branch approach with varying receptive fields—controlled by kernel size and dilation factors—to capture low-, middle-, and high-order spatial features simultaneously. It uses stacked $3 \times 3$ convolutions to mimic larger kernels efficiently.
  • Wave-based CMixer: Inspired by wave physics, this module treats feature channels as oscillating waves defined by amplitude and phase. By applying complex-valued weights and "wave superposition," the module reinforces important semantic features (in-phase) and suppresses noise (out-of-phase). This allows for a much smaller expansion ratio ($r=2$) while maintaining high performance.
The proposed Wave-based CMixer module utilizing wave superposition for channel aggregation.

Key Insights and Ideas

  • The "Middle-Order" Advantage: By explicitly targeting middle-order interactions, SpaRTAN captures the full semantics of an object. Activation maps show it recognizes entire structures (e.g., a whole panda) rather than just isolated features like eyes.
  • Dynamic Channel Allocation: Unlike static MLPs, the CMixer conceptualizes channels as waves, allowing the model to dynamically reallocate features based on the semantic context of the input image.
  • Hardware-Aware Design: The model uses full convolutions in early stages for GPU efficiency and switches to depthwise convolutions in later stages to optimize the accuracy-efficiency trade-off.

Key Results

SpaRTAN demonstrates remarkable parameter efficiency:

  • ImageNet-1k Classification: The SpaRTAN-T variant achieves 77.7% top-1 accuracy with only 3.8M parameters and 1.08 GFLOPs, rivaling models twice its size like ConvNeXt-XT (7.4M params).
  • Object Detection (COCO): Used as a backbone for RT-DETR, SpaRTAN-T achieved 50.0% AP, surpassing the ResNet-34 benchmark by 1.2% with 10 million fewer parameters.
  • Tiny Variant: The SpaRTAN-XT model reaches 74.4% accuracy with a mere 2.2M parameters, ideal for edge devices.

Limitations and Future Work

Future exploration includes scaling kernel sizes within this framework and assessing performance on downstream tasks like semantic segmentation and pose estimation. The authors also propose investigating model compression strategies like pruning and low-bit quantization for further efficiency.

Summary updated

BibTeX
@inproceedings{pay2025spartan,
  author = {Quan Bi Pay and Vishnu Monn Baskaran and Junn Yong Loo and KokSheik Wong and Simon See},
  title = {{SpaRTAN: Spatial Reinforcement Token-based Aggregation Network for Visual Recognition}},
  booktitle = {IJCNN},
  year = {2025},
  doi = {10.1109/ijcnn64981.2025.11228292},
}
Figure