text
stringlengths
0
820
Even as we adapted them, TTT andFixMatch do not
fully take advantage of the rich information obtained at de-
cision time in the VAS context as we proceed through each
input task: we not only observe the input image x, but
also observe query results over time during the search. We
therefore propose two new variants of TTA which are spe-
cific to the VAS setting: (a) Online TTA and (b) Stepwise
TTA. InOnline TTA , we update parameters of the policy
network after each task is completed during decision time,
which yields for us both the input xand the observations
oof the search results, which only partially correspond to
y, since we have only observed the contents of the previ-
ously queried grid cells. Nevertheless, we can simply use
this partial information oas a part of the REINFORCE pol-icy gradient update step to update the policy parameters θ.
InStepwise TTA , we update the policy network parameters,
even during the execution of a particular task, at decision
time, once every m<Csteps. The main difference between
Online andStepwise variations of our TTA approaches is
consequently the frequency of updates. Note that we can
readily compose both of these TTA approaches with con-
ventional TTA methods, such as TTT andFixMatch .
5. Experiments
Evaluation Metric We evaluate the proposed approaches
in terms of the average number of target objects discovered
(we shorten it to ANT ).
Baselines We compare the proposed V AS policy learning
framework with the following baselines:
1.random search , where each grid is chosen uniformly
at random among those which haven’t been explored,
2.greedy classification , in which we train a classifier ψgc
to predict whether a particular grid has a target object
and search the grids most likely to contain the target
until the search budget is exhausted, and
3.greedy selection , based on the approach by Uzkent and
Ermon [30] which trains a policy ψgswhich yields a
probability of zooming into each grid cell j. We select
grids according to ψgsuntil the budget Cis saturated.
4.active learning , in which we randomly select the first
grid to query and then choose C−1grids using a state-
of-the-art active learning approach by Yoo et al. [37].
5.conventional active search , an active search method by
Jiang et al. [15], using a low-dimensional feature rep-
resentation for each image grid from the same feature
extraction network as in our approach.
Query Costs We consider two ways of generating query
costs: (i) c(i, j)=1for all i, j, where Cis just the number
of queries, and (ii) c(i, j)is based on Manhattan distance
between iandj. Most of the results we present reflect the
second setting; the results for uniform query costs are qual-
itatively similar and provided in the Supplement.
Datasets We evaluate the proposed approach using two
datasets: xView [17] and DOTA [34]. xView is a satellite
imagery dataset which consists of large satellite images rep-
resenting 60 categories, with approximately 3000 pixels in
each dimensions. We use 67% and33% of the large satel-
lite images to train and test the policy network respectively.
DOTA is also a satellite imagery dataset. We re-scale the
original∼3000×3000px images to 1200 ×1200px. Un-
less otherwise specified, we use N=36non-overlapping
pixel grids each of size 200 ×200.
5.1. Results on the xView Dataset
We begin by evaluating the proposed approaches on the
xView dataset, varying search budgets C∈{25,50,75}and
5
number of grid cells N∈{30,48,99}. We consider two tar-
get classes: small car andbuilding . As the dataset contains
variable size images, take random crops of 2500×3000 for
N=30,2400×3200 pixels for N=48, and 2700×3300
forN=99, thereby ensuring equal grid cell sizes.
Table 1: ANT comparisons for the small car target class on xView.
Method C=25 C=50 C=75
random search (N=30) 3.41 3.95 4.52
greedy classification (N=30) 3.91 4.60 4.76
greedy selection [30] (N=30) 3.90 4.63 4.78
active learning [37] (N=30) 3.92 4.58 4.73
conventional active search [15] (N=30) 3.61 4.17 4.70
VAS (N=30) 4.61 7.49 9.88
random search (N=48) 3.20 3.66 4.11
greedy classification (N=48) 3.87 4.29 4.52
greedy selection [30] (N=48) 3.89 4.42 4.53
active learning [37] (N=48) 3.87 4.28 4.51
conventional active search [15] (N=48) 3.26 3.74 4.32
VAS (N=48) 4.56 7.45 9.63
random search (N=99) 1.10 2.15 2.96
greedy classification (N=99) 1.72 2.79 3.36
greedy selection [30] (N=99) 1.78 2.83 3.41
active learning [37] (N=99) 1.69 2.78 3.33
conventional active search [15] (N=99) 1.42 2.31 3.10
VAS (N=99) 2.72 4.42 5.78
Table 2: ANT comparisons for the building target class on xView.
Method C=25 C=50 C=75
random search (N=30) 3.97 4.94 5.39
greedy classification (N=30) 4.69 5.27 5.80
greedy selection [30] (N=30) 4.84 5.33 5.82
active learning [37] (N=30) 4.67 5.24 5.80
conventional active search [15] (N=30) 4.15 5.20 5.51
VAS (N=30) 5.65 9.31 12.20
random search (N=48) 3.47 3.96 4.26
greedy classification (N=48) 3.90 4.43 4.61
greedy selection [30] (N=48) 3.95 4.51 4.67