text
stringlengths
0
820
empirically show that active learning is inappropriate for
solving the active search problem.
However, current active search approaches typically lack
a pre-search training phase, and are therefore effective in
relatively low dimensions and for relatively simple model
classes such as k-nearest-neighbors. In VAS, in contrast,
our goal is to learn how to search , that is, to learn how to
best use information obtained from previous search queries
in choosing the next query. We experimentally demonstrate
the advantage of VASover conventional active search below.
Contributions We propose a deep reinforcement learning
approach to solve the VAS problem. Our key contribution
is a novel policy architecture which makes use of a natural
representation of search state, in addition to the task im-age input, which the policy uses to dynamically adapt to the
task at hand at decision time, without additional training.
Additionally, we consider a variant of VAS in which the na-
ture of input tasks at decision time is sufficiently different
to warrant test-time adaptation, and propose several adap-
tation approaches that take advantage of the VAS problem
structure. We extensively evaluate our proposed approaches
toVAS on two satellite imagery datasets in comparison with
several baseline, including a state-of-the-art approach for a
related problem of identifying regions of an image to zoom
into [30]. Our results show that our approach significantly
outperforms all baselines.
In summary, we make the following contributions:
• We propose visual active search (VAS) , a novel visual
reasoning model that represents an important class of
geospatial exploration problems, such as identifying
poaching activities, illegal trafficking, etc.
• We propose a deep reinforcement learning approach
forVAS that learns how to search for target objects in
a broad geospatial area based on aerial imagery.
• We propose two new variants of test-time adaptation
(TTA) variants of VAS: (a)Online TTA and (b) Stepwise
TTA, as well as an improvement of the FixMatch state-
of-the-art TTA method [26].
• We perform extensive experiments on two publicly
available satellite imagery datasets, xView and DOTA,
in a variety of settings, and demonstrate that proposed
approaches significantly outperform all baselines.
2. Related Work
Foveated Processing of Large Images Numerous pa-
pers [33, 31, 30, 36, 32, 22, 29, 20, 19, 35] have explored the
use of low-resolution imagery to guide the selection of im-
age regions to process at high resolution, including a num-
ber using reinforcement learning to this end. Our setting is
quite different, as we aim to choose a sequence of regions
to query, where each query yields the true label , rather than
a higher resolution image region, and these labels are im-
portant for both guiding further search, and as an end goal.
Reinforcement Learning for Visual Navigation Rein-
forcement learning has also been extensively used for vi-
sual navigation tasks, such as point and object localiza-
tion [5, 18, 21, 7]. While similar at the high level, these
tasks involve learning to decide on a sequence of visual
navigation steps based on a local view of the environment
and a kinematic model of motion, and commonly do not in-
volve search budget constraints. In our case, in contrast, the
full environment is observed initially (perhaps at low reso-
lution), and we sequentially decide which regions to query,
and are not limited to a particular kinematic model.
Active Search and Related Problem Settings Garnett et
al. [11] first introduced Active Search (AS) . Unlike Active
Learning [24], ASaims to discover members of valuable
2
and rare classes rather than on learning an accurate model.
Garnett et al. [11] demonstrated that for any l>m, al-step
lookahead policy can be arbitrarily superior than an m-step
one, showing that a nonmyopic active search approach can
be significantly better than a myopic one-step lookahead.
Jiang et al. [15, 14] proposed approaches for efficient non-
myopic active search, while Jiang et al. [13] introduced con-
sideration of search cost into the problem.
We note two crucial differences between our setting and
the previous works on active search. First, we are the first
to consider the problem in the context of vision, where the
problem is high-dimensional, while prior techniques rely on
a relatively low dimensional feature space. Second, we use
reinforcement learning as a means to learn a search policy,
in contrast to prior work on active search which aims to
design efficient search algorithms.
3. Model
At the center of our task is an aerial image xwhich is
partitioned into Ngrid cells, x=(x(1), x(2), ..., x(N)). We
can also view xas the disjoint union of these Ngrid cells,
each of which is a sub-image. A subset (possibly empty) of
these grid cells contain an instance of the target object. We
formalize this by associating each grid cell jwith a binary
label y(j)∈{0,1}, where y(j)=1iff grid cell jcontains
the target object. Let y=(y(1), y(2), ..., y(N)).
We do not know ya priori, but can sequentially query to
identify grid cells that contain the target object. Whenever
we query a grid cell j, we obtain both the associated label
y(j), (i.e., whether it contains the target object) andaccrue
utility if the queried cell actually contains a target object.
Our ultimate goal is to find as many target objects as possi-
ble through a sequence of such queries given a total query
budget constraint C,
Formally, let c(j, k)be the cost of querying grid cell kif