Joint self-attention and branch sampling for object detection on drone imagery

Yunzuo ZHANG; Cunyu WU; Yameng LIU; Tian ZHANG; Yuxin ZHENG

doi:10.37188/OPE.20233118.2723

Journals >Optics and Precision Engineering >Volume 31 >Issue 18 >Page 2723 > Article

Optics and Precision Engineering
Vol. 31, Issue 18, 2723 (2023)

Joint self-attention and branch sampling for object detection on drone imagery

Yunzuo ZHANG^1,2,*, Cunyu WU¹, Yameng LIU¹, Tian ZHANG¹, and Yuxin ZHENG¹

Author Affiliations

¹School of Information Science and Technology， Shijiazhuang Tiedao University， Shijiazhuang050043， China

²Hebei Key Laboratory of Electromagnetic Environmental Effects and Information Processing， Shijiazhuang Tiedao University， Shijiazhuang050043， China

show less

DOI: 10.37188/OPE.20233118.2723 Cite this Article

Yunzuo ZHANG, Cunyu WU, Yameng LIU, Tian ZHANG, Yuxin ZHENG. Joint self-attention and branch sampling for object detection on drone imagery[J]. Optics and Precision Engineering, 2023, 31(18): 2723 Copy Citation Text

show less

Fig. 1. Overall structure of the algorithm proposed in this paper

Download full size | View in the Article

Fig. 2. Structure diagram of NRCS

Download full size | View in the Article

Fig. 3. Structure diagram of DBS-FU

Download full size | View in the Article

Fig. 4. Structure diagram of FE-IR

Download full size | View in the Article

Fig. 5. Comparison of detection results of different algorithms on VisDrone2019

Download full size | View in the Article

Fig. 6. Comparison of detection results of different algorithms on UAVDT

Download full size | View in the Article

Method	指标	Pedestrian	People	Bicycle	Car	Van	Truck	Tricycle	Awning-tricycle	Bus	Motor	ALL
Baseline-s	mAP₅₀	66.8	53.7	35.8	88.4	56.3	45.2	42.4	21.5	64.6	64.3	53.9
Baseline-s	mAP	33.9	22.5	16.6	57.7	39.9	30.6	24.3	13.6	46.8	31.6	31.7
Ours-s	mAP₅₀	72.1	58.0	42.7	90.5	61.0	53.8	48.6	25.2	72.4	68.8	59.3
Ours-s	mAP	38.4	26.1	21.6	67.6	45.0	36.0	29.2	16.0	54.6	36.4	37.1
Baseline-m	mAP₅₀	69.4	56.0	39.7	89.4	59.1	52.6	46.8	26.3	71.4	66.0	57.7
Baseline-m	mAP	34.2	24.5	19.7	67.5	44.2	34.3	28.4	16.2	51.9	33.8	35.4
Ours-m	mAP₅₀	73.6	60.3	45.4	91.1	62.0	55.2	50.6	29.7	79.4	70.3	61.8
Ours-m	mAP	39.4	27.1	23.2	69.1	46.5	39.1	30.4	18.3	59.2	36.9	38.9

Table 1. Comparison of the results of the baseline and the algorithm in this paper on VisDrone2019

Method	mAP₅₀	mAP	mAP₇₅	Ap-small	AP-mid	AP-large
ClusDet^［10］	56.2	32.4	31.6	-	-	-
DSHNet^［11］	51.8	30.3	30.9	-	-	-
HRDNet^［12］	62.0	35.5	35.1	-	-	-
Yolov5-s	53.7	31.7	31.4	22.0	43.7	49.5
Yolov5-m	58.6	35.4	36.8	27.6	47.6	52.4
Yolov7^［22］	57.3	35.6	37.1	26.5	46.0	50.3
YoloX	53.5	31.4	31.7	22.5	41.5	48.5
MobileNetv3	55.4	32.9	32.9	24.5	44.3	49.5
MobileViT	55.5	33.3	33.7	24.9	44.2	41.8
mSODANet^［27］	55.9	36.9	37.4	-	-	-
Ours-s	59.3	37.1	38.3	29.2	48.5	51.6
Ours-m	62.1	38.9	39.5	31.2	51.3	52.5

Table 2. Comparison of results of different algorithm models on VisDrone2019

Method	指标	Car	Truck	Bus	All
Baseline-s	mAP₅₀	73.0	15.2	26.6	38.3
Baseline-s	mAP	41.2	8.1	15.9	21.7
Ours-s	mAP₅₀	73.3	19.4	39.7	44.1
Ours-s	mAP	40.9	10.9	23.0	24.9
Baseline-m	mAP₅₀	72.1	8.55	40.5	40.4
Baseline-m	mAP	39.6	4.86	22.9	22.5
Ours-m	mAP₅₀	73.7	18.0	40.9	44.2
Ours-m	mAP	42.3	11.2	24.1	25.8

Table 3. Comparison of the results of the baseline and the algorithm in this paper on UAVDT

Method	mAP₅₀	mAP	mAP₇₅
ClusDet^［10］	26.5	13.9	12.5
DMNet^［28］	24.6	14.7	16.3
GLSAN^［29］	30.5	19.0	21.7
CDMNet^［30］	29.1	16.8	18.5
DSHNet^［11］	30.4	17.8	19.7
Yolov5-s	38.3	21.7	22.6
Yolov5-m	41.4	22.6	24.1
PRDet^［31］	34.1	19.8	21.3
UFPMP-Det^［32］	38.7	24.6	28.0
Ours-s	44.1	24.9	27.3
Ours-m	44.7	25.3	28.1

Table 4. Comparison of results of different algorithm models on UAVDT

Method	Up-sampling			Down-sampling
Method	Nearest	Bilinear	DBUS	Conv	Pooling	DBDS
mAP₅₀	53.7	54.1	54.5	53.7	53.5	54.8
mAP	31.7	32.1	32.8	31.7	32.0	33.2

Table 5. Performance comparison of different sampling methods

	Baseline	Fourhead	Our-Fourhead	NRCS	DBUS	DBDS	FE-IR（3）	Parm	GFLOPs	mAP₅₀（%）	mAP（%）	FPS
1	√							7.028M	16.0	53.7	31.7	35.6
2	√	√						7.262M	19.3	54.8	32.4	24.7
3	√		√					6.911M	19.1	56.1	34.3	25.6
4	√		√	√				6.583M	19.3	56.9	35.1	27.1
5	√		√		√			6.942M	21.4	56.3	34.6	27.1
6	√		√			√		6.982M	21.6	56.8	35.2	26.9
7	√		√		√	√		7.013M	23.4	57.1	35.2	26.5
8	√		√	√	√	√		6.686M	23.6	58.3	35.9	26.9
9	√		√	√	√	√	√	7.875M	26.8	59.7	37.1	22.1

Table 6. Ablation experiment result

Yunzuo ZHANG, Cunyu WU, Yameng LIU, Tian ZHANG, Yuxin ZHENG. Joint self-attention and branch sampling for object detection on drone imagery[J]. Optics and Precision Engineering, 2023, 31(18): 2723

Download Citation

Tools

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享