Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
zoom-in to visualize the details of the pictures.
Figure 4: Left The histogram of number of views. Right the
histogram of standard deviation in years per area in fMoW
dataset.
studies utilizes geo-location of an image as a prior to im-
prove image recognition accuracy [33, 14, 24, 20, 6]. Other
studies [44, 11, 12, 42] use geo-tagged training datasets to
learn how to predict the geo-location of previously unseen
images at test time. In our study, we leverage geo-tag infor-
mation to improve unsupervised and self-supervised learn-
ing methods.
3. Problem Definition
We consider a geo-tagged visual dataset
{((x1
i,···,xTi
i),lati,loni)}N
i=1, where the ith data-point consists of a sequence of images Xi= (x1
i,···,xTi
i)
at a shared location, with latitude and longitude equal
to lati,lonirespectively, over time ti= 1,...,Ti. When
Ti>1, we refer to the dataset to have temporal information
or structure. Although temporal information is often not
available in natural image datasets ( e.g. ImageNet), it is
common in remote sensing. While the temporal structure is
similar to that of conventional videos, there are some key
differences that we exploit in this work. First, we consider
relatively short temporal sequences, where the time differ-
ence between two consecutive “frames” could range from
months to years. Additionally unlike conventional videos
we consider datasets where there is no viewpoint change
across the image sequence.
Given our setup, we want to obtain visual representa-
tionszti
iof imagesxti
isuch that the learned representation
can be transferred to various downstream tasks. We do not
assume access to any labels or human supervision beyond
the lati,lonigeo-tags. The quality of the representations
is measured by their performance on various downstream
tasks. Our primary goal is to improve the performance
of self-supervised learning by utilizing the geo-coordinates
and the unique temporal structure of remote sensing data.
3
3.1. Functional Map of the World
Functional Map of the World (fMoW) is a large-scale
publicly available remote sensing dataset [5] consisting of
approximately 363,571 training images and 53,041 test im-
ages across 62 highly granular class categories. It provides
images (temporal views) from the same location over time
(x1
i,···,xTi
i)as well as geo-location metadata (lati,loni)
for each image. Fig. 4 shows the histogram of the number
of temporal views in fMoW dataset. We can see that most of
the areas have multiple temporal views where Tican range
from 1 to 21, and on average there is about 2.5-3 years of
difference between the images from an area. Also, we show
examples of spatially aligned images in Fig. 2. As seen in
Fig. 5, fMoW is a global dataset consisting of images from
seven continents which can be ideal for learning global re-
mote sensing representations.
3.2. GeoImageNet
Following [7], we extract geo-coordinates for a sub-
set of images in ImageNet [8] using the FLICKR API.
More specifically, we searched for geo-tagged images
in ImageNet using the FLICKR API, and were able to
find 543,435 images with their associated coordinates
(lati,loni)across 5150 class categories. This dataset is
more challenging than ImageNet-1k as it is highly imbal-
anced and contains about 5 ×more classes. In the rest of
the paper, we refer to this geo-tagged subset of ImageNet as
GeoImageNet .
We show some examples from GeoImageNet in Fig. 3.
As shown in the figure, for some images we have geo-
coordinates that can be predicted from visual cues. For ex-
ample, we see that a picture of a person with a Sombrero
hat was captured in Mexico. Similarly, an Indian Elephant
picture was captured in India, where there is a large popula-
tion of Indian Elephants. Next to it, we show the picture of
an African Elephant (which is larger in size). If a model is
trained to predict where in the world the image was taken,
it should be able to identify visual cues that are transferable
to other tasks (e.g., visual cues to differentiate Indian Ele-
phants from the African counterparts). Figure 5 shows the
distribution of images in the GeoImageNet dataset.
4. Method
In this section, we briefly review contrastive loss func-
tions for unsupervised learning and detail our proposed ap-
proach to improve Moco-v2 [3], a recent contrastive learn-
ing framework, on geo-located data.
4.1. Contrastive Learning Framework
Contrastive [13, 2, 3, 34, 27] methods attempt to learn
a mappingfq:xt
i↦→zt
i∈Rdfrom raw pixels xt
ito
semantically meaningful representations zt
iin an unsuper-
Figure 5: Topshows the distribution of the fMoW and Bot-