Split (2)

train · 64k rows

text stringlengths 0 820
on aerial images. To this end, we train different networks from
scratch with the help of the largest RS scene recognition dataset
up to now — MillionAID, to obtain a series of RS pretrained
backbones, including both convolutional neural networks (CNN)
and vision transformers such as Swin and ViTAE, which have
shown promising performance on computer vision tasks. Then,
we investigate the impact of RSP on representative downstream
tasks including scene recognition, semantic segmentation, object
detection, and change detection using these CNN and vision
transformer backbones. Empirical study shows that RSP can
help deliver distinctive performances in scene recognition tasks
and in perceiving RS related semantics such as “Bridge” and
“Airplane”. We also ﬁnd that, although RSP mitigates the data
discrepancies of traditional ImageNet pretraining on RS images,
it may still suffer from task discrepancies, where downstream
tasks require different representations from scene recognition
tasks. These ﬁndings call for further research efforts on both
large-scale pretraining datasets and effective pretraining meth-
ods. The codes and pretrained models will be released at
https://github.com/ViTAE-Transformer/RSP.
Index Terms —Remote Sening Pretraining, CNN, Vision Trans-
former, Classiﬁcation, Detection, Semantic Segmentation.
I. I NTRODUCTION
WITH the development of geoinformatics technology,
the earth observation ﬁelds have witnessed signiﬁcant
progress, where various remote sensing (RS) sensors and
devices have been widely used. Among them, with the advan-
tages of real-time, abundant amount, and easy access, the aerial
image has become one of the most important data sources in
earth vision to serve the requirements of a series of practical
tasks, such as precision agriculture [1], [2] and environmental
monitoring [3]. In these applications, aerial scene recognition
is a fundamental and active research topic over the past years.
However, because of the own characteristics of aerial images,
it is still challenging to efﬁciently understand the aerial scene.
The aerial images are usually obtained by a camera in a bird-
view perspective lying on the planes or satellites, perceiving a
D. Wang, B. Du and Gui-Song Xia are with the School of
Computer Science, Wuhan University, Wuhan 430072, China (e-mail:
wd74108520@gmail.com; dubo@whu.edu.cn; guisong.xia@whu.edu.cn). B.
Du is the corresponding author.
J. Zhang is with the School of Computer Science, Faculty of Engineering,
The University of Sydney, Australia (jing.zhang1@sydney.edu.au).
D. Tao is with the JD Explore Academy, China and is also with the School
of Computer Science, Faculty of Engineering, The University of Sydney,
Australia (dacheng.tao@gmail.com).
Playground
Swimming pool
Tennis court
Building
(c)
(a)
(d)(b)
Playground
PlaygroundFig. 1. The challenges of aerial scene recognition. (a) and (b) are the natural
image and aerial image belonging to the “park” category. (c) and (d) are two
aerial images from the “school” category. Despite the distinct view difference
between (a) and (b), (b) contains the playground that is unusual in the park
scenes but usually exists in the school scenes like (d). On the other hand,
(c) and (d) show different colors as well as signiﬁcantly different spatial
distributions of land objects like the playground and swimming pool. Here, (a)
is obtained from http://travel.qunar.com/p-oi24486013-townhill country park
by searching “park” on internet, while (b), (c), and (d) are the aerial images
from the AID dataset.
large scope of land uses and land covers. The obtained aerial
scene is usually difﬁcult to be interpreted since the interference
of the scene-irrelevant regions and the complicated spatial
distribution of land objects. Speciﬁcally, it causes the issue of
inter-class similarity in aerial scene understanding, i.e., some
different scenes present similar characteristics, as well as the
issue of large intra-class variations, where the scenes in the
same category have discrepancies, as shown in Figure 1.
To tackle the above problems, it is necessary to obtain dis-
criminative feature representations for different categories of
aerial scenes. According to the difference in feature extraction
methods, they can be divided into three types, i.e., the hand-
crafted features, the unsupervised learning features, and the
supervised deep learning (DL) features. Initially, researchers
directly utilize simple properties, such as color [4], texture
[5], contour [6], spectral or their combination [7] to recognize
different aerial scenes. Besides these intuitive attributes, there
are also some well-designed feature descriptors. For instance,
the scale-invariant feature transformation and histogram of
oriented gradients. These handcrafted features usually perform
well in simple scenes while being ineffective in complex
scenes. They are usually regarded as shallow features from
a modern view in the DL era, while interpreting complex
scenes requires more semantic information, which can not be
efﬁciently extracted by shallow-layer methods [8]. Compared
with the above approaches, the unsupervised learning methodsarXiv:2204.02825v4 [cs.CV] 4 May 2023
2 JOURNAL OF L ATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
provide a feasible way to automatically extract the suitable
features by adaptively learning the mapping functions or ﬁlters
based on a set of handcrafted features or the raw pixel intensity
values. The typical unsupervised learning methods include
potential latent semantic analysis [9] and bag-of-visual-words
[10]. Some simple feature enhancement methods such as the
principal component analysis also belong to this category.
Nonetheless, the encoded unsupervised features still have
limited performance since no category supervision is explicitly

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.