text
stringlengths
0
820
map fine-grained textual descriptions (captions). Our ap-
proach allows us to theoretically map anything that can be
expressed in natural language, and thus, serves as a general
framework for zero-shot mapping.
Recently, several works [7, 8, 9] have delved into model-
ing the relationship between images and text. Models such
as CLIP [7] and ALBEF [8] are trained on large database of
captioned-images to learn a multimodal embedding space
that unifies vision and text space. These embedding spaces
can be utilized to learn the textual descriptors of a given
image. However, a limitation of using overhead imagery
is its tendency to provide coarse and generic features of a
given area. We observe that this property of overhead im-
ages holds true in the CLIP embedding space as well, where
these images are related to coarse textual concepts like city,
beach, or property. These images capture a broad perspec-
tive from above, offering limited insight into the intricate
concepts and dynamics within the location. Ground-level
images, on the other hand, provide more detailed infor-
mation about a place. The CLIP embedding space has a
better understanding of fine-grained concepts for ground-
level images since it was primarily trained on them and
their descriptive captions. Yet, several challenges hinder
the direct utilization of ground-level imagery for mapping
tasks. Firstly, ground-level images are sparsely available
i.e., obtaining a ground-level image for every location on
Earth is not feasible. Secondly, the coverage and quality of
a ground-level image for the same location can have large
variations which could introduce unwanted variations dur-
ing inference.
To address these issues, we present a novel weakly-
supervised cross-view approach for learning fine-grained,
and dynamic textual concepts for geographic locations.
First, we create a large-scale dataset with paired overhead
and ground-level images. Our dataset uses a subset of the
1arXiv:2307.15904v1 [cs.CV] 29 Jul 2023
Figure 1: Captions generated by the CLIPCAP model [1] using CLIP embeddings vs. Dynamic Sat2Cap embeddings. (Row-
1) shows the results from CLIP embeddings which produce many generic descriptions. (Row-2 and 3) shows the results from
our Sat2Cap embeddings for the month of May and January respectively. The captions generated using Sat2Cap embeddings
are more fine-grained and dynamic. While (d) does not add any winter properties for the January query, this behavior is
expected as the image is over Australia where January falls in the middle of summer.
YFCC100M [10]. More details about the dataset are pre-
sented in Section 3.1. Using this paired dataset, we learn the
CLIP distribution of the ground-level scene for a given loca-
tion. CLIP embeddings of ground-level images can describe
detailed textual concepts of that location. Our Sat2Cap
model learns to predict the expected CLIP embedding of the
ground-level scene using the overhead image. Compared to
the CLIP embeddings, Sat2Cap embeddings tend to capture
more fine-grained textual concepts for a given geolocation
as seen in Figure 1.
To account for the temporal associations between vari-
ous concepts and a location, our model is conditioned on
temporal data, specifically, the date and time stamps from
the Flickr imagery. This allows our model to learn fine-
grained concepts that can be dynamically adapted to dif-
ferent date and time settings. Figure 1 shows an example
of CLIP-generated coarse captions vs. Sat2Cap-generated
fine-grained dynamic captions.
Our method is also weakly-supervised and thus does not
require any text-labeled data. Creating a large-scale dataset
of fine-grained captions and geolocation can be challeng-
ing. However, our approach only requires geotagged and
timestamped ground-level images which are easily accessi-
ble and scalable. Additionally, our framework is designed
to learn high-resolution information of a location. This rich
information can be used as an additional signal to solve a
number of other downstream tasks. The following pointssummarize the primary contributions of our work:
• A novel weakly-supervised approach for learning fine-
grained dynamic textual concepts of geographic loca-
tions
• A model for effective cross-view image retrieval be-
tween overhead images and ground-level images taken
in the wild
• A zero-shot approach for creating large-scale textual
maps
• A new large-scale cross-view dataset
2. Related Works
2.1. Deep Learning Based Mapping
Creating maps of attributes of interest is an important
task in many domains. Deep Learning methods have been
used extensively [11, 12, 13, 14, 15] in recent years to make
mapping efficient and scalable. Alhassan et al. [16] fine-
tuned imagenet pretrained models to make landcover pre-
dictions. Similarly [17, 18] leveraged large-scale annotated
data from different sensors to improve landuse and land-
cover classification using deep learning methods. Apart
from remote sensing, other areas have also leveraged deep
learning for their own mapping tasks. Using high-resolution
2
images, [19] trained several deep learning architectures for
automated CNN-based mapping of Martian rockfalls. On
the other hand, [20] used an unsupervised approach to map
regions with high “Au” deposits.
There have been other works that specifically focus on
mapping visual attributes. For instance, [21] used features
from both overhead and ground-level imagery, and intro-
duced a cross-view approach to map scenicness. Later
works focused on creating dynamic maps. Both [22, 2]
conditioned their model on temporal information along with