text
stringlengths
0
820
module for the large vehicle target class on DOTA.
Method C=25 C=50 C=75
VAS-DINO (N=36) 4.56 6.75 8.03
VAS-ViT (N=36) 4.60 6.82 8.09
VAS (N=36) 4.63 6.79 8.07
VAS-DINO (N=64) 5.27 8.44 10.45
VAS-ViT (N=64) 5.31 8.51 10.48
VAS (N=64) 5.33 8.47 10.51
E. More Visual Illustration of V AS and the
Most Competitive Greedy Selection base-
line Method
In this section, we provide additional visualization of
comparative exploration behaviour of V AS and and the most
competitive greedy selection baseline approach. In figure 7,
we compare the search strategy with large vehicle as a tar-
get class. In figure 8, we compare the behaviour with small
caras a target class. In figure 9, we analyze the exploration
behaviour with ship as a target class.
These additional visualizations again justify the efficacy
of V AS over the strongest baseline method.
Figure 7: Comparison of policies learned using VAS(left) and the greedy
selection baseline method (right).
Figure 8: Comparison of policies learned using VAS(left) and the greedy
selection baseline method (right).
Figure 9: Comparison of policies learned using VAS(left) and the greedy
selection baseline method (right).
F. Assessment of V AS and other Competitive
Baseline Methods with a Different Evalua-
tion Metric
We additionally compare the search performance of V AS
with all the baseline methods using a metric, which we call
Effective Success Rate (ESR) . A na ¨ıve way to evaluate the
proposed approaches is to simply use success rate , which is
is the fraction of total search steps Kthat identify a target
object. However, if Kexceeds the total number of target
objects in x, normalizing by Kis unreasonable, as even a
perfect search strategy would appear to work poorly. Con-
sequently, we propose effective success rate (ESR) as the
efficacy metric, defined as follows:
ESR=#Targets Discovered
min{#Targets , K }(4)
Thus, we divide by the number of targets one can possibly
discover given a search budget K, rather than simply the
search budget.
F.1. Results on the xView Dataset with ESR as Eval-
uation Metric
We initiate our analysis by assessing the proposed
methodologies using the xView dataset, for varying search
13
budgets K∈{12,15,18}and number of grid cells N∈
{30,48,99}. We also consider two target classes for our
search: 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 for
N=99, thereby guarantees uniform grid cell dimensions
across the board.
Table 16: ESR comparisons for the small car target class
on the xView dataset.
Method K=12 K=15 K=18
random search (N=30) 0.598 0.632 0.704
greedy classification (N=30) 0.619 0.675 0.718
greedy selection [30] (N=30) 0.627 0.684 0.729
VAS (N=30) 0.766 0.826 0.861
random search (N=48) 0.489 0.517 0.558
greedy classification (N=48) 0.512 0.551 0.589
greedy selection [30] (N=48) 0.524 0.568 0.596
VAS (N=48) 0.694 0.722 0.741
random search (N=99) 0.336 0.369 0.378
greedy classification (N=99) 0.365 0.384 0.405
greedy selection [30] (N=99) 0.376 0.395 0.418
VAS (N=99) 0.564 0.587 0.602
Table 17: ESR comparisons for the building target class on
the xView dataset.
Method K=12 K=15 K=18
random search (N=30) 0.663 0.681 0.697
greedy classification (N=30) 0.701 0.734 0.767
greedy selection [30] (N=30) 0.708 0.740 0.786
VAS (N=30) 0.854 0.886 0.912
random search (N=48) 0.526 0.547 0.556
greedy classification (N=48) 0.548 0.569 0.585
greedy selection [30] (N=48) 0.552 0.574 0.604
VAS (N=48) 0.677 0.716 0.738
random search (N=99) 0.443 0.462 0.483
greedy classification (N=99) 0.460 0.482 0.504
greedy selection [30] 0.469 0.488 0.514
VAS (N=99) 0.654 0.676 0.690
The results are presented in Table 16 for the small car
class and in Table 17 for the building class. We see sig-
nificant improvements in performance of the proposed VAS
approach compared to all baselines, ranging from ∼15–50%
improvement relative to the most competitive state-of-the-
art method, greedy selection .F.2. Results on the DOTA Dataset with ESR as Eval-
uation Metric
We also conduct our experiments on the DOTA dataset.
We use large vehicle andship as our target classes. In both
cases, we also report results with non-overlapping pixel
grids of size 200×200and150×150(N=36andN=64,
respectively). We again use K∈{12,15,18}.
Table 18: ESR comparisons for the large vehicle target
class on the DOTA dataset.