text
stringlengths
0
820
overhead images to learn dynamic concepts for a given loca-
tion. While the prior works mostly focus on mapping some
specific attribute, we attempt to generalize the mapping pro-
cess. Hence, in our work, we introduce a framework to cre-
ate maps of free-form textual prompts, which we call Tex-
tual Maps.
2.2. Vision-Language Pretraining
Recently, Vision-Language (VL) models have shown
great promise in their ability to model complex relation-
ships between the vision and text space. ConVIRT [23]
and VirTex [24] both introduced methods that used image-
text pairs to learn rich visual representations. CLIP [7]
demonstrated the results of VL pretraining on a large-scale
dataset (400M pairs) and validated the efficacy of large-
scale VL pretraining for several downstream tasks. Flo-
rence [25] and ALIGN [26] further increased the scale of
data by training on 900M and 1.8B pairs respectively. Other
works [8, 9, 27, 28] have since focused on learning better
VL embedding space. With the existence of these pow-
erful pretrained VL models, many researchers have uti-
lized their embedding spaces to solve specific downstream
tasks. CLIPCap [1], and [29] used CLIP space to gener-
ate image captions. Other models like [30, 31, 32] uti-
lized the CLIP space for text-to-image generation. Several
works [33, 34, 35] have also used the CLIP space for im-
age retrieval tasks. In our work, we utilize the rich CLIP
space to bridge the gap between geolocations and their fine-
grained textual descriptions.
3. Method
Our objective is to learn an embedding space that
describes the expected ground-level scene given a geo-
graphic location and an overhead image. Secondly, our
embedding space needs to dynamically adapt to temporal
manipulations for the same location. We have ground-
level images {g1, g2, ...gn}, corresponding overhead im-
ages{o1, o2, ...on}, and respective metadata for the ground-
level images {e1, e2, ...en}. Each eicontains the latitude
and longitude information of the sample, as well as the date
and time when the ground-level image was captured. We
also have a CLIP image encoder fθthat generates CLIP em-
beddings for a given ground-level image.3.1. Dataset
We created a large-scale cross-view dataset to train our
model. The ground-level images in the dataset are taken
from the YFCC100M [10] dataset. The YFCC100M dataset
contains 99.3 million images, collected from Flickr. Our
cross-view dataset uses a smaller sample from this collec-
tion which excludes all US imagery. Our dataset contains
close to 6M images. Each of these images has a geoloca-
tion, timestamp, and other meta information such as tags,
description, camera type, etc. For each Flickr image, we
download an overhead image centered at its location. We
use the Bing Maps API to download 800x800patch satel-
lite images at 0.6m/px resolution.
3.2. Approach
We initialize our Sat2Cap image encoder gθwith
the weights of fθ. A batch of ground-level images
{g1, g2, ..., g k}is passed through the CLIP encoder to get
the ground-level CLIP embeddings. These embeddings
serve as the target for alignment. A batch of correspond-
ing overhead images {o1, o2, ..., o k}is passed through the
Sat2Cap image encoder to obtain the embeddings, as fol-
lows:
Gi=fθ(gi) (1)
Oi=gθ(oi) (2)
To align the overhead image embeddings with the
ground-level CLIP embeddings, we contrastively train our
model using the InfoNCE [36] loss as follows:
L=1
kk∑
i=0−logexp(Oi·Gi/τ)
∑k
j=0exp(Oi·Gj/τ)(3)
We optimize this loss to minimize the distance between
co-located overhead and ground-level images in the CLIP
space. It is worth noting that throughout the training pro-
cess, the CLIP image encoder remains frozen. Hence,
with our training procedure, we essentially allow the over-
head images to move close to images from their respective
ground-level scene in the CLIP space. Our results from
Section 4.1 show that Sat2Cap learns a strong correlation
between co-located overhead and ground-level images.
3.3. Learning Dynamic Concepts of Places
Many ground-level concepts are temporally dependent.
Concepts like ‘crowded street’, ‘snowy place’ etc can dra-
matically vary based on the exact time we query about them.
In order to model such dynamic concepts, we condition
Sat2Cap on the timestamps of the ground-level images.
For each sample, we extract the year, month, day, and
hour in which the ground-level image was taken. We also
add the geolocation information to provide a stronger signal
3
Figure 2: A weakly-supervised learning framework to learn fine-grained and dynamic concepts of geolocations without
explicit text labels.
Method Overhead2Ground (10K) Ground2Overhead (10K)
Model Dynamic Encoder Dropout Meta Information R@5 ↑R@10 ↑Median-R ↓ R@5 ↑R@10 ↑Median-R ↓
CLIP - - - 0.007 0.013 1700 0.108 0.019 2857
ours✗ ✗ ✗ 0.398 0.493 15 0.356 0.450 11
✓ ✗ ✗ 0.322 0.413 34 0.254 0.343 20
✓ ✗ ✓ 0.368 0.467 23 0.298 0.398 13
✓ ✓ ✗ 0.467 0.564 13.5 0.366 0.462 7