Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
Many problems can be viewed as forms of geospatial
search aided by aerial imagery, with examples ranging from
detecting poaching activity to human trafficking. We model
this class of problems in a visual active search (VAS) frame-
work, which has three key inputs: (1) an image of the entire
search area, which is subdivided into regions, (2) a local
search function, which determines whether a previously un-
seen object class is present in a given region, and (3) a
fixed search budget, which limits the number of times the
local search function can be evaluated. The goal is to max-
imize the number of objects found within the search budget.
We propose a reinforcement learning approach for VAS that
learns a meta-search policy from a collection of fully anno-
tated search tasks. This meta-search policy is then used to
dynamically search for a novel target-object class, lever-
aging the outcome of any previous queries to determine
where to query next. Through extensive experiments on sev-
eral large-scale satellite imagery datasets, we show that the
proposed approach significantly outperforms several strong
baselines. We also propose novel domain adaptation tech-
niques that improve the policy at decision time when there
is a significant domain gap with the training data. Code is
publicly available at this link.
1. Introduction
Consider a large national park that hosts endangered an-
imals, which are also in high demand on a black market,
creating a major poaching problem. An important strategy
in an anti-poaching portfolio is to obtain aerial imagery us-
ing drones that helps detect poaching activity, either ongo-
ing, or in the form of traps laid on the ground [1, 2, 3, 9, 8].
The quality of the resulting photographs, however, is gen-
erally somewhat poor, making the detection problem ex-
tremely difficult. Moreover, park rangers can only inspect
relatively few small regions to confirm poaching activity,
doing so sequentially. Crucially, inspecting such regionsyields new ground truth information about poaching activ-
ity that we can use to decide which regions to inspect in
the future. We can distill some key generalizable structure
Figure 1: A comparison of a greedy search policy (dashed line)
with an active search strategy (solid line) for the small car tar-
get class. The greedy policy is not able to adapt when a car is
not found in the starting cell and needlessly searches many similar
cells. The active strategy adapts and explores regions with differ-
ent visual characteristics, eventually finding the objects of interest.
from this scenario: given a broad area image (often with
a relatively low resolution), sequentially query small areas
within it (e.g., by sending park rangers to the associated re-
gions, on the ground), with each query returning the ground
truth, to identify as many target objects (e.g., poachers or
traps) as possible. The number of queries we can make is
typically limited, for example, by budget or resource con-
straints. Moreover, query results (e.g., detected poaching
activity in a particular region) are highly informative about
the locations of target objects in other regions , for example,
due to spatial correlation. We refer to this general modeling
framework as visual active search (VAS) . Numerous other
scenarios share this broad structure, such as identification
of drug or human trafficking sites, broad area search-and-
rescue, identifying landmarks, and many others.
A simple solution to the broad-area search problem is to
divide the broad area into many grid cells, train a classifier
1arXiv:2211.15788v3 [cs.CV] 29 Oct 2023
to predict existence of a target object in each grid cell, and
simply explore the top Kcells in terms of predicted likeli-
hood of the object being in the cell. We call this the greedy
policy, which essentially reduces geospatial active search to
the familiar object identification (or detection) problem. In
Figure 1, we offer some intuition about why this idea fails
to capture important problem structure. Suppose we look
for small cars in an image, starting in the grid marked start,
which we initially think is the best place to look. The greedy
policy being a one-shot predictor of likely grid cells con-
taining target, continues to explore similar regions (marked
as). What this approach ignores, as does framing the
problem as traditional one-shot object identification, is the
fact that both success and failure of past queries are infor-
mative due to complex spatial correlation among objects
and other patterns in the scene; here, because the car was
not found, we proceed to instead explore regions that have
somewhat different characteristics. The key to visual active
search, therefore, is to learn how to make use of the ground
truth information obtained over a sequence of queries to de-
cide where to query next. Additionally, Section 5 of [11]
provides a rigorous analysis of the general sub-optimality
of greedy policies in active search settings.
Relationship to Active Search and Active Learning VAS
is closely related to active search [11, 10, 15, 13]. Ac-
tive search is typically concerned with binary classification,
and aims to maximize the number of discovered positively-
labeled inputs. It uses a function f, which predicts labels of
inputs x, as a means to this end, with each query serving the
dual-purpose of improving fas well as identifying a posi-
tive instance. A central concern in active search, therefore,
is achieving a balance of exploration (learning f) and ex-
ploitation (identifying target inputs). This consideration is
also the key distinction between active search and active
learning [25], which is concerned solely with improving
the predictive quality of f. Thus, if we had only a single
query, active learning would typically choose xfor which
prediction is highly uncertain, whereas active search would
choose xfor which f(x)is most confidently positive. We