Sparse Transformer Based Remote Sensing Rotated Object Detection

Linyuan He; Junqiang Bai; Xu He; Chen Wang; Xulun Liu

doi:10.3788/LOP202259.1810003

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 18 >Page 1810003 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 18, 1810003 (2022)

Sparse Transformer Based Remote Sensing Rotated Object Detection

Linyuan He^1、2、*, Junqiang Bai¹, Xu He², Chen Wang², and Xulun Liu²

Author Affiliations

¹Unbanned System Research Institute, Northwestern Polytechnical University, Xi’an , Shaanxi 710072, China

²School of Aeronautical Engineering, Air Force Engineering University, Xi’an , Shaanxi 710038, China

show less

DOI: 10.3788/LOP202259.1810003 Cite this Article Set citation alerts

Linyuan He, Junqiang Bai, Xu He, Chen Wang, Xulun Liu. Sparse Transformer Based Remote Sensing Rotated Object Detection[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1810003 Copy Citation Text

show less

Fig. 1. Structure diagram of proposed model

Download full size

Fig. 2. Schematic diagrams of multiple self attention. (a) Neighborhood related self attention; (b) cross step self attention; (c) sparse self attention

Download full size

Fig. 3. Geometric representation of rotated bounding box

Download full size

Fig. 4. Detection results of proposed model on different datasets. (a) DOTA; (b) UCAS-AOD

Download full size

Fig. 5. Visualization of test results of different models. (a) RoI-Trans model; (b)

O^{2}

DETR model; (c) proposed model

Download full size

Model	R2CNN^［12］	RRPN^［13］	CAD-Net^［14］	RoI-Trans^［15］	$O^{2}$ DETR^［8］	Proposed model
Pl	80.89	88.52	87.80	88.64	86.01	89.91
BD	65.75	71.20	82.40	78.52	75.92	85.78
BR	35.34	31.66	49.40	43.44	46.02	50.65
GTF	67.44	59.30	73.50	75.92	66.65	78.16
SV	59.93	51.85	71.10	68.81	79.70	64.34
LV	50.91	56.19	63.50	73.68	79.93	75.43
SH	55.81	57.25	76.70	83.59	89.17	75.78
TC	90.67	90.81	90.90	90.74	90.44	90.88
BC	66.92	72.84	79.20	77.27	81.19	78.67
ST	72.39	67.38	73.30	81.46	76.00	84.45
SBF	55.06	56.69	48.40	58.39	56.91	57.91
RA	52.23	52.84	60.90	53.54	62.45	63.56
HA	55.14	53.08	62.00	62.83	64.22	64.56
SP	53.35	51.94	67.00	58.93	65.80	66.74
HC	48.22	53.58	62.20	47.67	58.96	66.33

Table 1. AP values of different models on DOTA dataset

Model	mAP /%	Backbone	Frame rate /（frames·s^-1）
RRPN^［13］	61.01	VGG-16	5.25
R2CNN^［12］	60.67	VGG-16	3.81
CAD-Net^［14］	69.90	ResNet101	5.82
RoI-Trans^［15］	69.56	ResNet101	5.76
$O^{2}$ DETR^［8］	72.15	ResNet101	6.58
Proposed model	72.87	ResNet101	6.78

Table 2. mAP values and detection frame rate of different detection algorithms on DOTA dataset

Model	Plane	Car	mAP
RRPN^［13］	88.04	74.36	81.20
R2CNN^［12］	89.76	78.89	84.32
R-DFPN^［16］	88.91	81.27	85.09
P-RSDet^［11］	92.69	87.38	90.03
Proposed model	91.22	89.58	90.40

Table 3. AP values of different models on UCAS-AOD dataset

Linyuan He, Junqiang Bai, Xu He, Chen Wang, Xulun Liu. Sparse Transformer Based Remote Sensing Rotated Object Detection[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1810003

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information