Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
step 7
step 9
step 11
step 13
step 15
(b)(Top row )Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS. (Middle row )
Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS while enforcing the query
outcomes at every stage being “ unsuccessful ”.(Bottom row )Query sequences, and corresponding heat maps (darker indicates higher
probability), obtained using V AS while enforcing the query outcomes at every stage being “ successful ”.
Figure 13: Sensitivity Analysis of VAS with a sample test image and large vehicle as target class under distance based query
cost.
19
(a) The original image
step 1
step 3
step 5
step 7
step 9
step 11
step 13
step 15
(b)(Top row )Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS. (Middle row )
Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS while enforcing the query
outcomes at every stage being “ unsuccessful ”.(Bottom row )Query sequences, and corresponding heat maps (darker indicates higher
probability), obtained using V AS while enforcing the query outcomes at every stage being “ successful ”.
Figure 14: Sensitivity Analysis of VAS with a sample test image and caras target class under distance based query cost.
20
(a) The original image
step 1
step 3
step 5
step 7
step 9
step 11
step 13
step 15
(b)(Top row )Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS. (Middle row )
Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS while enforcing the query
outcomes at every stage being “ unsuccessful ”.(Bottom row )Query sequences, and corresponding heat maps (darker indicates higher
probability), obtained using V AS while enforcing the query outcomes at every stage being “ successful ”.
Figure 15: Sensitivity Analysis of VAS with a sample test image and ship as target class under distance based query cost.
21
(a) The original image with query sequence.
step 1 step 5 step 10 step 15
(b) Saliency maps (red indicates high saliency), obtained using V AS at different stages of search process with large vehicle as target.
Figure 16: Saliency map visualization of VAS under uniform cost budget.
22
(a) The original image with query sequence.
step 1 step 5 step 10 step 15
(b) Saliency maps (red indicates high saliency), obtained using V AS at different stages of search process with small car as target.
Figure 17: Saliency map visualization of VAS under uniform cost budget.
23
(a) The original image with query sequence.
step 1 step 5 step 10 step 15
(b) Saliency maps (red indicates high saliency), obtained using V AS at different stages of search process with small car as target.
Figure 18: Saliency map visualization of VAS under uniform cost budget.
24
Sat2Cap: Mapping Fine-Grained Textual Descriptions from Satellite Images
Aayush Dhakal1Adeel Ahmad1,2Subash Khanal1Srikumar Sastry1Nathan Jacobs1
1Washington University in St. Louis2Taylor Geospatial Institute
Abstract
We propose a novel weakly supervised approach for cre-
ating maps using free-form textual descriptions (or cap-
tions). We refer to this new line of work of creating tex-
tual maps as zero-shot mapping. Prior works have ap-
proached mapping tasks by developing models that predict
over a fixed set of attributes using overhead imagery. How-
ever, these models are very restrictive as they can only solve
highly specific tasks for which they were trained. Map-
ping text, on the other hand, allows us to solve a large va-
riety of mapping problems with minimal restrictions. To
achieve this, we train a contrastive learning framework
called Sat2Cap on a new large-scale dataset of paired over-
head and ground-level images. For a given location, our
model predicts the expected CLIP embedding of the ground-
level scenery. Sat2Cap is also conditioned on temporal in-
formation, enabling it to learn dynamic concepts that vary
over time. Our experimental results demonstrate that our
models successfully capture fine-grained concepts and ef-
fectively adapt to temporal variations. Our approach does
not require any text-labeled data making the training eas-
ily scalable. The code, dataset, and models will be made
publicly available.
1. Introduction
Creating maps of different attributes is an important task
in many domains. Traditionally, methods of mapping in-
volve exhaustive data collection across vast regions, which
is both time-consuming and labor-intensive. To address this
issue, recent studies have explored the use of Deep Learn-
ing models, with their strong visual learning capabilities, to
directly predict attributes of interest through overhead im-
agery. Salem et al. [2] used overhead images to map tran-
sient attributes [3] and scene categories [4] across large re-
gions, while Streltsov et al. [5] predicted residential build-
ing energy consumption using overhead imagery. Similarly,
Bency et al. [6] also used satellite images to map housingprices. However, all these prior methods focused on learn-
ing some specific pre-defined attributes. These attribute-
specific models are quite restrictive as they cannot map any-
thing beyond their preset list of variables. To overcome this
limitation, we created a novel framework that enables us to