Dataset for Oriented Ship Recognition (DOSR)

This dataset has been published in our paper Fine-Grained Recognition for Oriented Ship Against Complex Scenes in Optical Remote in Remote Sensing Images, doi:10.1109/TGRS.2021.3123666. Related ship recognition code is published on here

I. Introduction

DOSR is a public available dataset for fine-grained ship recognition with 1066 optical remote sensing images and 6127 ship instances. Images in DOSR are collected from Google Earth. The image size is ranged from 600 pixels to 1300 pixels with resolution of 0.5m-2.5m. To expand the diversity of data, we collect images from quantities of countries, including the United States, China, France, Japan, etc. The shooting time spans from 2001 to 2020.

In DOSR, there are 832 nearshore scenes containing 5212 instances and 234 offshore scenes containing 915 instances. The dataset is divided into three sets: training set with 523 images, validation set with 223 images and testing set with 320 images. Fig.1 displays examples of 20 fine-grained classes in DOSR. Instance amount of each class is shown in Table I.

Fig. 1 Examples of 20 Fine-grained Classes in DOSR

Table I Instance Amount

Class Name	Class Short Name	Total amount	Train amount	Val amount	Train+Val amount	Test amount
Transpot	Tra.	1682	820	358	1178	504
Yacht	Yac.	1188	583	214	797	391
Speedboat	Spe.	529	166	69	235	294
Auxiliary Ship	Aux.	426	216	95	311	115
Military Ship	Mli.	419	204	68	272	147
Tug	Tug	293	169	58	227	66
Fishing Boat	Fis.	276	166	49	215	61
Bulk Cargo Vessel	BCV.	275	144	60	204	71
Cargo	Car.	249	109	51	160	89
Container	Con.	177	92	38	130	47
Cruise	Cru.	165	86	30	116	49
Deckbarge	DeB.	100	51	20	71	29
Tanker	Tan.	86	45	20	65	21
Deckship	DeS.	69	40	12	52	17
Flat Traffic Ship	FTS.	59	42	6	48	11
Floating Crane	Flo.	38	16	13	29	9
Multihull	Mul.	36	14	12	26	10
Barge	Bar.	34	23	4	27	7
Communication Ship	Com.	14	5	3	8	6
Submarine	Sub.	12	7	1	8	4

Fig.2 shows several typical scenes in DOSR such as small scene, small and dense scene, regular dense scene, irregular dense scene, scale variance scene, and clutter scene.

Fig. 2 Typical Scenes in DOSR

II. Annotations

We provide 2 versions of annotations for study.

1. Five-parameter Annotations

All images are annotated in VOC format. Each instance is represented by a five-parameter tuple (x,y,w,h,θ), which is consisted with our paper "Fine-Grained Recognition for Oriented Ship Against Complex Scenes in Optical Remote Sensing Images". Fig.3 explains the meaning of five parameters.

Fig. 3 Parameters of oriented bounding box. (a) Reference rectangle, θ=0；(b) θ=(0,2/pi); (c) θ=pi; (d) (2/pi,pi)

2. Eight-parameter Annotations

All images are annotated in VOC format. Each instance is represented by an eight-parameter tuple (x1,y1,x2,y2,x3,y3,x4,y4), which describes the coordinates of the four vertices of the oriented bounding box. In addition, we also provide annotations of horizontal bounding boxes in this version. We obtained horizontal bounding boxes by calculating minimum bounding rectangles of oriented bounding boxes, and use 4-parameter (xmin,ymin,xmax,ymax) to describe the horizontal bounding boxes.

III. How to obtain

You can get DOSR at (1) Google Drive or (2) Baidu Netdisk with access code 1596