MAN and CAT: Mix Attention to NN and Concatenate Attention to YOLO

===This is the official page of Mix Attention, MANet and CAT-YOLO===

Benchmark Results

Mix Attention Block with ResNet-101 and WRN-18 on CIFAR-10

Model	Params(M)	Top-1 Error(%)
ResNet-101 + Mix Attention	50.07	6.17
WRN-18 + Mix Attention	27.11	4.77

Mix Attention Block with ResNet-101 and WRN-18 on CIFAR-100

Model	Params(M)	Top-1 Error(%)
ResNet-101 + Mix Attention	50.07	23.19
WRN-18 + Mix Attention	27.11	19.11

---

MANet on ImageNet

Model	Params(M)	Top-1 Accuracy(%)
MANet-B	69.3	81.7
MANet-S	23.4	78.3
MANet-T	4.3	73.1

MANet on CIFAR-10

Model	Params(M)	Top-1 Accuracy(%)
MANet-B	69.3	97.2
MANet-S	23.4	95.1
MANet-T	4.3	93.4

MANet on CIFAR-100

Model	Params(M)	Top-1 Accuracy(%)
MANet-B	69.3	88.7
MANet-S	23.4	86.5
MANet-T	4.3	81.6

---

CAT-YOLO on COCO 2017

Model	Backbone	Params(M)	Latency(ms)	AP
CAT-YOLO-v1	CSPDarknet53-Tiny	6.16	9.9(TITAN RTX)	24.1
CAT-YOLO-v2	MANet-T	9.17	12.7(TITAN RTX)	25.7
CAT-YOLO-v3	MANet-T	12.5	16.8(TITAN RTX)	33.5

---

User Guide

MAN includes the plug-and-play modules(Mix Attention) and backbone(MANet).

Note:

1. We divide the modules and backbones for CIFAR and ImageNet respectively.

2. The sub-folder named "Big Version" in Modules play the role of one individual layer.

3. The sub-folder named "Tiny Version" in Modules play the role of the enhance module in the network's bottleneck.

CAT includes the files of CAT-YOLO.

Citation

  @article{guan2022man,
  title={MAN and CAT: mix attention to nn and concatenate attention to YOLO},
  author={Guan, Runwei and Man, Ka Lok and Zhao, Haocheng and Zhang, Ruixiao and Yao, Shanliang and Smith, Jeremy and Lim, Eng Gee and Yue, Yutao},
  journal={The Journal of Supercomputing},
  pages={1--29},
  year={2022},
  publisher={Springer}
  }