IEEE Geoscience and Remote Sensing Magazine - December 2023 - 24

prior knowledge with abundant base-class samples, and the
few-shot fine-tuning stage leverages the prior knowledge to
facilitate the learning of few-shot novel concepts. The research
on remote sensing FSOD mainly focuses on metalearning
methods [254], [255], [256], [257], [258], [259] and
transfer learning methods [260], [261], [262], [263], [264],
[265], [266], [267], [268], [269].
The metalearning-based methods acquire task-level
knowledge by simulating a series of few-shot learning
tasks and generalize this knowledge to tackle the few-shot
learning of novel classes. Li et al. [255] first employed metalearning
for remote sensing FSOD and achieved satisfactory
detection performance with only one to 10 labeled
samples. Later, a series of metalearning-based few-shot detectors
were developed in the remote sensing community
[254], [255], [256], [257], [258], [259]. For example, Cheng
et al. [254] proposed a prototype CNN to generate better
foreground proposals and class-aware ROI features for remote
sensing FSOD by learning class-specific prototypes.
Wang et al. [258] presented a metametric training paradigm
to enable a few-shot learner with flexible scalability for fast
adaptation to few-shot novel tasks.
Transfer learning-based methods aim at fine-tuning
common knowledge learned from abundant annotated
data to few-shot novel data and typically consist of a base
training stage and a few-shot fine-tuning stage. Huang et al.
[266] proposed a balanced fine-tuning strategy to alleviate
the number imbalance problem between novel-class samples
and base-class samples. Zhou et al. [265] introduced
proposal-level contrast learning in the fine-tuning phase to
learn more robust feature representations in few-shot scenarios.
Compared with the metalearning-based methods,
the transfer learning-based method has a simpler and more
memory-efficient training paradigm.
DATASETS AND EVALUATION METRICS
DATASETS INTRODUCTION AND SELECTION
Datasets have played an indispensable role throughout
the development of object detection in RSIs. On the one
hand, datasets serve as common ground for the performance
evaluation and comparison of detectors. On the
other hand, datasets push researchers to address increasingly
challenging problems in the RSOD field. In the past
TABLE 3. COMPARISONS OF WIDELY USED DATASETS IN THE FIELD OF RSOD.
DATASET
SOURCE
TAS [270]
NWPU VHR-10 [18]
VEDAI [272]
DLR 3K [273]
UCAS-AOD [274]
COWC [275]
HRSC [276]
RSOD [43]
SSDD [277]
LEVIR [278]
xView [2]
DOTA v1.0 [117]
HRRSD [48]
DIOR [28]
Google Earth
SZTAKI-INRIA [271] QuickBird, Ikonos, and
Google Earth
Google Earth
Utah Automated Geographic
Reference Center
DLR 3K camera system
Google Earth
Multiple sources
Google Earth
Google Earth and Tianditu
RadarSat-2, TerraSAR-X, and
Sentinel-1
Google Earth
WorldView-3
Google Earth, Jilin-1, and
Gaofen-2
1
1
ANNOTATION CATEGORIES INSTANCES IMAGES IMAGE WIDTH RESOLUTION YEAR
HBB
OBB
1,319
665
HBB
OBB
OBB
OBB
Point
OBB
HBB
HBB
HBB
HBB
Google Earth and Baidu Maps HBB
Google Earth
AIR-SARShip-1.0 [279] Gaofen-3
MAR20 [280]
FGSD [281]
DOSR [282]
AI-TOD [283]
FAIR1M [34]
DOTA-v2.0 [33]
SODA-A [284]
HBB
HBB
Google Earth
Google Earth
Google Earth
Multiple sources
Gaofen satellites and
Google Earth
Google Earth, Jilin-1, Gaofen-2,
and airborne images
Google Earth
OBB: oriented bounding box.
*Average image width.
24
10
9
8
2
1
26
4
1
3
60
HBB and OBB 15
13
20
1
HBB and OBB 20
OBB
OBB
HBB
OBB
43
20
8
37
HBB and OBB 18
OBB
9
3,651
2,950
14,235
6,029
32,716
2,976
6,950
2,456
11,000
1 million
188,282
55,740
190,288
3,000
22,341
5,634
6,172
700,621
1,020,579
1,793,658
800,203
30
9
800
1,268
20
910
53
1,061
976
1,160
792
~800
~1,000
1,024
5,616
~1,000
-
0.5-1 m
0.3-2 m
0.125 m
0.13 m
0.3-2 m
2,000-19,000 0.15 m
~1,100
~1,000
500
22,000 800-600
1,413
2,806
~3,000
800-13,000
21,761
152-10,569
23,463 800
31
3,824
2,612
1,066
3,000
~800
930
600-1,300
28,036 800
42,796
11,268
2,510
600-10,000
800-20,000
0.4-2 m
0.3-3 m
1-15 m
0.2-1 m
0.3 m
0.1-1 m
2008
2012
2014
2015
2015
2015
2016
2016
2017
2017
2018
2018
2018
0.15-1.2 m 2019
0.5-30 m 2019
1 and 3 m 2019
2020
-
0.12-1.93 m 2020
0.5-2.5 m 2021
2021
-
0.3-0.8 m 2021
0.1-4.5 m 2021
4,761 × 2,777* -
2022
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2023
IEEE Geoscience and Remote Sensing Magazine - December 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2023
