[ WACV24]SBCFormer：快速、轻量级 神经网络模型 Efficient CNN-ViT hybrid neural network backbone

一： SBCFormer [ WACV Paper], [Code]

SBCFormer 是一种 CNN-ViT 轻量混合网络backbone，能在单板计算机上能以每秒 1 帧速度进行全尺寸 ImageNet 分类的轻量级网络。
在低端 CPU 上实现高准确性和快速计算，如在树莓派 4-B ARM-Cortex A72 CPU 上提供了以往无法达到的每帧1秒运行速度。

SBCFormer: Lightweight Network Capable of Full-size ImageNet Classification at 1 FPS on Single Board Computers

• This paper introduces a CNN-ViT hybrid network called SBCFormer, which achieves high accuracy and fast computation on such low-end CPUs.
• We compare our SBCFormers against a wide range of relevant and up-todate alternatives.
• SBCFormer uses the proposed hourglass attention computation to aggregate global information from the entire image while minimizing computational costs.
• SBCFormer achieves the highest trade-off between accuracy and speed on a Raspberry Pi 4 Model B with an ARM-Cortex A72 CPU.

二：网络架构

三、性能

ImageNet-1K Classification

COCO 2017 Object Detection

四、Reference

@inproceedings{lu2024sbcformer,
  title={SBCFormer: Lightweight Network Capable of Full-size ImageNet Classification at 1 FPS on Single Board Computers},
  author={Lu, Xiangyong and Suganuma, Masanori and Okatani, Takayuki},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={1123--1133},
  year={2024}
}