RFAConv: Innovating Spatial Attention and Standard Convolutional Operation(preprint)

This repository is a PyTorch implementation of our paper: RFAConv: Innovating Spatial Attention and Standard Convolutional Operation.

If you are interested in our other work, you can find information on https://github.com/CV-ZhangXin/LDConv.

The code has been opened, thank you for your support.

In the classification experiment, the code for Resnet comes from https://github.com/zgcr/pytorch-ImageNet-CIFAR-COCO-VOC-training

In the detection experiment, the YOLOv5&YOLOv8 code comes from https://github.com/ultralytics/yolov5, and the YOLOv7 code comes from https://github.com/WongKinYiu/yolov7.

The pre-training weights of the model on ImageNet-1k can be found in https://pan.baidu.com/s/1kvPyLie0OeOB2CGQxdUT0g. The verification code is "rfac".

---

Numerical results on ImageNet

Mdels	FLOPS(G)	Params(M)	Top1	Top5
Resnet18	1.82	11.69	69.59	89.05
+CAMConv(r)	1.83	11.75	70.76	89.74
+CBAMConv(r)	1.83	11.75	69.38	89.12
+CAConv(r)	1.83	11.74	70.58	89.59
+RFAConv(r)	1.91	11.85	71.23	90.29
+RFCAConv(r)	1.92	11.89	72.01	90.64
+RFCBAMConv(r)	1.90	11.88	72.15	90.71

Mdels	FLOPS(G)	Params(M)	Top1	Top5
Resnet34	3.68	21.80	73.33	91.37
+CAMConv(r)	3.68	21.93	74.03	91.69
+CBAMConv(r)	3.68	21.93	72.95	91.26
+CAConv(r)	3.68	21.91	73.76	91.68
+RFAConv(r)	3.84	22.16	74.25	92.03

Object detection

Results on Pascal VOC 2007+2012 validation sets

Mdels	FLOPS(G)	Params(M)	mAP50	mAP	time
YOLOv5n	4.2	1.8	67.8	41.5	2.7
+CAMConv(r)	4.2	1.8	67.8	41.4	2.9
+CBAMConv(r)	4.3	1.8	68.1	41.9	3.0
+CAConv(r)	4.3	1.8	68.4	42.4	3.0
+RFAConv(r)	4.5	1.8	69.5	43.3	3.0

Mdels	FLOPS(G)	Params(M)	mAP50	mAP	time
YOLOv5s	15.9	7.1	74.4	48.9	3
+CAMConv(r)	16.0	7.1	73.9	48.5	3.5
+CBAMConv(r)	16.0	7.1	74.1	49.0	3.7
+CAConv(r)	16.1	7.1	75	49.6	3.1
+RFAConv(r)	16.4	7.2	75	50.0	5.1
+RFCBAMConv(r)	16.4	7.2	75.1	50.1	3.9
+RFCAConv(r)	16.6	7.2	75.6	51.0	4.4

Mdels	FLOPS(G)	Params(M)	mAP50	mAP	time
YOLOv7-tiny	13.2	6.1	76.4	50.2	5.0
+CAMConv(r)	13.2	6.1	76.3	50.3	5.4
+CBAMConv(r)	13.2	6.1	76.5	50.1	5.4
+CAConv(r)	13.2	6.1	76.6	50.5	5.4
+RFAConv(r)	13.6	6.1	76.7	50.6	7.5

Mdels	FLOPS(G)	Params(M)	mAP50	mAP	time
YOLOv8n	8.1	3.0	74.0	53.5	3.0
+CAMConv(r)	8.1	3.0	73.8	52.8	3.1
+CBAMConv(r)	8.2	3.0	74.4	53.3	3.1
+CAConv(r)	8.2	3.0	74.5	53.8	2.9
+RFAConv(r)	8.4	3.1	74.7	54.0	3.2

Results on COCO validation sets

Mdels	FLOPS(G)	Params(M)	AP50	AP75	AP	APs	APm	APl	time
YOLOv5n	4.5	1.8	45.6	28.9	27.5	13.5	31.5	35.9	4.4
+CAMConv(r)	4.5	1.8	45.6	28.3	27.4	13.8	31.4	35.8	5.2
+CBAMConv(r)	4.5	1.8	45.5	28.6	27.6	13.6	31.2	36.6	5.4
+CAConv(r)	4.5	1.8	46.2	29.2	28.1	14.3	32	36.6	4.8
+RFAConv(r)	4.7	1.9	47.3	30.6	29.0	14.8	33.4	37.4	5.3

 

Mdels	FLOPS(G)	Params(M)	AP50	AP75	AP	APs	APm	APl	time
YOLOv7-tiny	13.7	6.2	53.8	38.3	35.9	19.9	39.4	48.8	6.8
+RFAConv(r)	14.1	6.3	55.1	40.1	37.1	20.9	41.1	50.0	8.4

 

Mdels	FLOPS(G)	Params(M)	AP50	AP75	AP	APs	APm	APl	time
YOLOv8n	8.7	3.1	51.9	39.7	36.4	18.4	40.1	52	4.2
+CAMConv(r)	8.8	3.1	51.6	39.0	36.2	18.0	39.9	51.2	4.5
+CBAMConv(r)	8.8	3.1	51.5	39.6	36.3	18.3	40.1	51.5	4.6
+CAConv(r)	8.8	3.1	52.1	39.9	36.7	17.8	40.3	51.6	4.3
+RFAConv(r)	9.0	3.2	53.4	41.1	37.7	18.9	41.8	52.7	4.5
+RFCAConv(r)	9.1	3.2	53.9	41.7	38.2	19.7	42.3	53.5	4.7

 

Citation

You may want to cite:

@misc{zhang2023rfaconv,
      title={RFAConv: Innovating Spatial Attention and Standard Convolutional Operation}, 
      author={Xin Zhang and Chen Liu and Degang Yang and Tingting Song and Yichen Ye and Ke Li and Yingze Song},
      year={2023},
      eprint={2304.03198},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}