HRSC2016_SOTA

This repo collects some state-of-the-art results on remote sensing ship dataset HRSC2016 for convenient comparison.

(Ranked according to the mAP(07)).

model	backbone	input_size	mAP(07)	mAP(12)	FPS	paper link	code
R2CNN	ResNet101	800x800	73.07	79.73	2	arxiv	tf
RC1&RC2	VGG16	-	75.7	-	-	ICPRAM	-
Axis Learning	ResNet101	800x800	78.15	-	14	Remote Sensing	-
RRPN	ResNet101	800x800	79.08	85.64	3.5	TMM	-
TOSO	ResNet101	800x800	79.29	-	17	ICASSP2020	-
RRD	VGG16	800x800	84.30	-	-	CVPR2018	C++
RoI-Transformer	ResNet101	512x800	86.20	-	6	CVPR2019	pytorch
RSDet	ResNet50	800x800	86.5	-	-	AAAI2021	tf
Gliding Vertex	ResNet101	512×800	88.20	-	10	TPAMI	pytorch
OPLD	ResNet50	1024x1333	88.44	-	7.3	JSTAR	pytorch
BBAVectors	ResNet101	608x608	88.60	-	11.7	WACV2021	pytorch
DRN	Hourglass104	768x768	-	92.70	10	CVPR2020	pytorch
DAL	ResNet101	416x416	88.95	-	34	AAAI2021	pytorch
RIDet-Q	ResNet101	800x800	89.10	-	8.5	arxiv	pytorch
R3Det	ResNet101	800x800	89.26	96.01	12	AAAI2021	tf, pytorch
DCL	ResNet101	800x800	89.46	96.41	-	CVPR2021	tf
SLA	ResNet101	768x768	89.51	-	-	Remote Sensing	pytorch
CSL	ResNet50	800x800	89.62	96.10	-	ECCV2020	tf
RIDet-O	ResNet101	800x800	89.63	-	-	arxiv	pytorch
CFC-Net	ResNet101	800x800	89.70	-	28	TGRS	pytorch
GWD	ResNet101	800x800	89.85	97.37	-	ICML2021	tf
S2ANet	ResNet101	512x800	90.17	95.01	12.7	TGRS	pytorch
ReDet	ResNet101	512x800	90.46	97.63	-	CVPR2021	pytorch
Oriented R-CNN	ResNet101	1333x800	90.50	97.60	15.1	ICCV2021	pytorch

Notes

For fair comparison, experiments on the HRSC2016 dataset should meet the following requirements:

Strictly following the official dataset division. Specifically, training and testing should be performed on the training set and test set respectively. It is unreasonable to use the validation set for training.
Many previous work used the VOC07 method to calculate mAP, and some recent work used the VOC12 standard to calculate mAP. These two results should not be compared equally.
All results are obtained at level 1, that is, there is only one class (ship).