gremlin97/RemoteSensingCorpus · Datasets at Hugging Face

text stringlengths 0 820
tom shows the distribution of GeoImageNet.
vised way. The training objective encourages representa-
tions corresponding to pairs of images that are known a pri-
ori to be semantically similar (positive pairs) to be closer
to each other than typical unrelated pairs (negative pairs).
With similarity measured by dot product, recent approaches
in contrastive learning differ in the type of contrastive loss
and generation of positive and negative pairs. In this work,
we focus on the state-of-the-art contrastive learning frame-
work MoCo-v2 [3], an improved version of MoCo [13], and
study improved methods for the construction of positive and
negative pairs tailored to remote sensing applications.
The contrastive loss function used in the MoCo-v2
framework is InfoNCE [27], which is deﬁned as follows for
a given data sample:
Lz=−logexp(z·ˆz/λ)
exp(z·ˆz/λ) +∑N
j=1exp(z·kj/λ),(1)
wherezandˆzare query and key representations obtained
by passing the two augmented views of xt
i(denotedvand
v′in Fig. 1) through query and key encoders, fqandfkpa-
rameterized by θqandθkrespectively. Here zandˆzform
a positive pair. The Nnegative samples, {kj}N
j=1, come
from a dictionary of representations built as a queue. We
refer readers to [13] for details on this. λ∈R+is the tem-
perature hyperparameter.
The key idea here is to encourage representations of pos-
itive (semantically similar) pairs to be closer, and negative
4
(semantically unrelated) pairs to be far apart as measured
by dot product. The construction of positive and negative
pairs plays a crucial role in this contrastive learning frame-
work. MoCo and MoCo-v2 both use perturbations (also
called “data augmentation”) from the same image to create
a positive example and perturbations from different images
to create a negative example. Commonly used perturbations
include random color jittering, random horizontal ﬂip, and
random grayscale conversion.
2016-04-17T15:49:27Z2012-11-21T15:17:29Z2016-11-10T16:00:51Z2016-11-10T16:00:51Z2011-06-06T15:56:51Z
Figure 6: Demonstration of temporal positives in eq. 2. An
image from an area is paired to the other images includ-
ing itself from the same area captured at different time. We
show the time stamps for each image underneath the im-
ages. We can see the color changes in the stadium seatings
and surrounding areas.
Temporal Positive Pairs Different from many commonly
seen natural image datasets, remote sensing datasets of-
ten have extra temporal information, meaning that for a
given location (lati,loni), there exists a sequence of spa-
tially aligned images Xi= (x1
i,···,xTi
i)over time. Unlike
in traditional videos where nearby frames could experience
large changes in content ( e.g. from a cat to a tree), in re-
mote sensing the content is often more stable across time
due to the ﬁxed viewpoint. For instance, a place on ocean
is likely to remain as ocean for months or years, in which
case satellite images taken across time at the same location
should share high semantic similarities. Even for locations
where non-trivial changes do occur over time, certain se-
mantic similarities could still remain. For instance, key fea-
tures of a construction site are likely to remain the same
even as the appearance changes due to seasonality.
Given these observations, it is natural to leverage tempo-
ral information for remote sensing while constructing pos-
itive or negative pairs since it can provide us with extra
semantically meaningful information of a place over time.
More speciﬁcally, given an image xt1
icollected at time t1,
we can randomly select another image xt2
ithat is spatially
aligned with xt1
i(i.e.xt2
i∈Xi). We then apply perturba-
tions ( e.g. random color jittering) as used in MoCo-v2 to the
spatially aligned image pair xt1
iandxt2
i, providing us with
atemporal positive pair (denotedvandv′in Figure 1) that
can be used for training the contrastive learning frameworkby passing them through query and key encoders, fqandfk
respectively (see Fig. 1). Note that when t1=t2, the tem-
poral positive pair is the same as the positive pair used in
MoCo-v2.
Given a data sample xt1
i, our TemporalInfoNCE objec-
tive function can be formulated as follows:
Lzt1
i=−logexp(zt1
i·zt2
i/λ)
exp(zt1
i·zt2
i/λ) +N∑
j=1exp(zt1
i·kj/λ),(2)
wherezt1
iandzt2
iare the encoded representations of the

tom shows the distribution of GeoImageNet.

vised way. The training objective encourages representa-

tions corresponding to pairs of images that are known a pri-

ori to be semantically similar (positive pairs) to be closer

to each other than typical unrelated pairs (negative pairs).

With similarity measured by dot product, recent approaches

in contrastive learning differ in the type of contrastive loss

and generation of positive and negative pairs. In this work,

we focus on the state-of-the-art contrastive learning frame-

work MoCo-v2 [3], an improved version of MoCo [13], and

study improved methods for the construction of positive and

negative pairs tailored to remote sensing applications.

The contrastive loss function used in the MoCo-v2

framework is InfoNCE [27], which is deﬁned as follows for

a given data sample:

Lz=−logexp(z·ˆz/λ)

exp(z·ˆz/λ) +∑N

j=1exp(z·kj/λ),(1)

wherezandˆzare query and key representations obtained

by passing the two augmented views of xt

i(denotedvand

v′in Fig. 1) through query and key encoders, fqandfkpa-

rameterized by θqandθkrespectively. Here zandˆzform

a positive pair. The Nnegative samples, {kj}N

j=1, come

from a dictionary of representations built as a queue. We

refer readers to [13] for details on this. λ∈R+is the tem-

perature hyperparameter.

The key idea here is to encourage representations of pos-

itive (semantically similar) pairs to be closer, and negative

4

(semantically unrelated) pairs to be far apart as measured

by dot product. The construction of positive and negative

pairs plays a crucial role in this contrastive learning frame-

work. MoCo and MoCo-v2 both use perturbations (also

called “data augmentation”) from the same image to create

a positive example and perturbations from different images

to create a negative example. Commonly used perturbations

include random color jittering, random horizontal ﬂip, and

random grayscale conversion.

2016-04-17T15:49:27Z2012-11-21T15:17:29Z2016-11-10T16:00:51Z2016-11-10T16:00:51Z2011-06-06T15:56:51Z

Figure 6: Demonstration of temporal positives in eq. 2. An

image from an area is paired to the other images includ-

ing itself from the same area captured at different time. We

show the time stamps for each image underneath the im-

ages. We can see the color changes in the stadium seatings

and surrounding areas.

Temporal Positive Pairs Different from many commonly

seen natural image datasets, remote sensing datasets of-

ten have extra temporal information, meaning that for a

given location (lati,loni), there exists a sequence of spa-

tially aligned images Xi= (x1

i,···,xTi

i)over time. Unlike

in traditional videos where nearby frames could experience

large changes in content ( e.g. from a cat to a tree), in re-

mote sensing the content is often more stable across time

due to the ﬁxed viewpoint. For instance, a place on ocean

is likely to remain as ocean for months or years, in which

case satellite images taken across time at the same location

should share high semantic similarities. Even for locations

where non-trivial changes do occur over time, certain se-

mantic similarities could still remain. For instance, key fea-

tures of a construction site are likely to remain the same

even as the appearance changes due to seasonality.

Given these observations, it is natural to leverage tempo-

ral information for remote sensing while constructing pos-

itive or negative pairs since it can provide us with extra

semantically meaningful information of a place over time.

More speciﬁcally, given an image xt1

icollected at time t1,

we can randomly select another image xt2

ithat is spatially

aligned with xt1

i(i.e.xt2

i∈Xi). We then apply perturba-

tions ( e.g. random color jittering) as used in MoCo-v2 to the

spatially aligned image pair xt1

iandxt2

i, providing us with

atemporal positive pair (denotedvandv′in Figure 1) that

can be used for training the contrastive learning frameworkby passing them through query and key encoders, fqandfk

respectively (see Fig. 1). Note that when t1=t2, the tem-

poral positive pair is the same as the positive pair used in

MoCo-v2.

Given a data sample xt1

i, our TemporalInfoNCE objec-

tive function can be formulated as follows:

Lzt1

i=−logexp(zt1

i·zt2

i/λ)

exp(zt1

i·zt2

i/λ) +N∑

j=1exp(zt1

i·kj/λ),(2)

wherezt1

iandzt2

iare the encoded representations of the