text stringlengths 0 820 |
|---|
application to geo-located datasets, e.g. remote sensing, |
where unlabeled data is often abundant but labeled data |
is scarce. We first show that due to their different char- |
acteristics, a non-trivial gap persists between contrastive |
and supervised learning on standard benchmarks. To close |
the gap, we propose novel training methods that exploit the |
spatio-temporal structure of remote sensing data. We lever- |
age spatially aligned images over time to construct tempo- |
ral positive pairs in contrastive learning and geo-location |
to design pre-text tasks. Our experiments show that our |
proposed method closes the gap between contrastive and |
supervised learning on image classification, object detec- |
tion and semantic segmentation for remote sensing. More- |
over, we demonstrate that the proposed method can also be |
applied to geo-tagged ImageNet images, improving down- |
stream performance on various tasks. Project Webpage can |
be found at this link geography-aware-ssl.github.io. |
1. Introduction |
Inspired by the success of self-supervised learning meth- |
ods [3, 13], we explore their application to large-scale re- |
mote sensing datasets (satellite images) and geo-tagged nat- |
ural image datasets. It has been recently shown that self- |
supervised learning methods perform comparably well or |
even better than their supervised learning counterpart on im- |
age classification, object detection, and semantic segmenta- |
tion on traditional computer vision datasets [21, 10, 13, 3, |
2]. However, their application to remote sensing images is |
largely unexplored, despite the fact that collecting and la- |
*Equal Contribution. Contact: {kayush, buzkent, chen- |
lin}@cs.stanford.edubeling remote sensing images is particularly costly as anno- |
tations often require domain expertise [37, 38, 36, 16, 5]. |
In this direction, we first experimentally evaluate the per- |
formance of an existing self-supervised contrastive learning |
method, MoCo-v2 [13], on remote sensing datasets, finding |
a performance gap with supervised learning using labels. |
For instance, on the Functional Map of the World (fMoW) |
image classification benchmark [5], we observe an 8% gap |
in top 1 accuracy between supervised and self-supervised |
methods. |
To bridge this gap, we propose geography-aware con- |
trastive learning to leverage the spatio-temporal structure |
of remote sensing data. In contrast to typical computer vi- |
sion images, remote sensing data are often geo-located and |
might provide multiple images of the same location over |
time. Contrastive methods encourage closeness of represen- |
tations of images that are likely to be semantically similar |
(positive pairs). Unlike contrastive learning for traditional |
computer vision images where different views (augmenta- |
tions) of the same image serve as a positive pair, we pro- |
pose to use temporal positive pairs from spatially aligned |
images over time. Utilizing such information allows the |
representations to be invariant to subtle variations over time |
(e.g., due to seasonality). This can in turn result in more |
discriminative features for tasks focusing on spatial vari- |
ation, such as object detection or semantic segmentation |
(but not necessarily for tasks involving temporal variation |
such as change detection). In addition, we design a novel |
unsupervised learning method that leverages geo-location |
information, i.e., knowledge about where the images were |
taken. More specifically, we consider the pretext task of |
predicting where in the world an image comes from, similar |
to [11, 12]. This can complement the information-theoretic |
objectives typically used by self-supervised learning meth- |
ods by encouraging representations that reflect geograph- |
ical information, which is often useful in remote sensing |
tasks [31]. Finally, we integrate the two proposed methods |
1arXiv:2011.09980v7 [cs.CV] 8 Mar 2022 |
Figure 1: Left shows the original MoCo-v2 [3] framework. Right shows the schematic overview of our approach. |
into a single geography-aware contrastive learning objec- |
tive. |
Our experiments on the functional Map of the World [5] |
dataset consisting of high spatial resolution satellite im- |
ages show that we improve MoCo-v2 baseline significantly. |
In particular, we can improve the accuracy on target ap- |
plications utilizing image recognition [5], object detec- |
tion [39, 1], and semantic segmentation [46]. In particular, |
we improve it by∼8%classification accuracy when testing |
the learned representations on image classification, ∼2% |
AP on object detection, ∼1%mIoU on semantic segmen- |
tation, and∼3%top-1 accuracy on land cover classifica- |
tion ˙Interestingly, our geography-aware learning can even |
outperform the supervised learning counterpart on temporal |
data classification by ∼2%. To further demonstrate the ef- |
fectiveness of our geography-aware learning approach, we |
extract the geo-location information of ImageNet images |
using FLICKR API similar to [7], which provides us with |
a subset of 543,435 geo-tagged ImageNet images. We ex- |
tend the proposed approaches to geo-located ImageNet, and |
show that geography-aware learning can improve the per- |
formance of MoCo-v2 by ∼2%on image classification, |
showing the potential application of our approach to any |
geo-tagged dataset. Figure 1 shows our contributions in de- |
tail. |
2. Related Work |
Self-supervised methods use unlabeled data to learn rep- |
resentations that are transferable to downstream tasks ( e.g. |
image classification). Two commonly seen self-supervised |
methods are pre-text task andcontrastive learning . |
Pre-text tasks Pre-text task based learning [22, 41, 29, 49, |
43, 28] can be used to learn feature representations when |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.