text
stringlengths
0
820
Illustration of V AS and the Most Competitive Greedy Se-
lection baseline Method; (F) Assessment of V AS and other
Baseline Methods with a Different Evaluation Metric; (G)
Search Performance Comparison with Other Policy Learn-
ing Algorithm (PPO); (H) Sensitivity Analysis of VAS; (I)
Efficacy of TTA on Search Tasks involving Large Number
of Grids; (J) Saliency map visualization of VAS;
A. Performance of V AS under Uniform Query
Cost
In this section, we report the performance of V AS under
uniform query cost. The results are presented in the follow-
ing Table 8 for the small car target class and in Table 9 for
thebuilding class from the xView dataset. We observe sig-
nificant improvements in performance of the proposed V AS
approach compared to all baselines, ranging from 11−25%
improvement relative to the most competitive greedy selec-
tionapproach.
Table 8: ANT comparisons for the small car target class.
Method C=12 C=15 C=18
random search (N=30) 4.57 5.66 6.85
greedy classification (N=30) 5.31 6.24 7.25
greedy selection [30] (N=30) 5.47 6.45 7.46
active learning [37] (N=30) 5.28 6.21 7.22
conventional AS [15] (N=30) 4.86 5.97 6.92
VAS (N=30) 6.03 7.24 8.24
random search (N=48) 3.80 4.97 5.98
greedy classification (N=48) 4.69 5.48 6.79
greedy selection [30] (N=48) 4.92 5.81 6.98
active learning [37] (N=48) 4.68 5.46 6.78
conventional AS [15] (N=48) 3.96 5.45 6.14
VAS (N=48) 5.62 6.81 7.86
random search (N=99) 3.12 3.61 4.45
greedy classification (N=99) 3.68 4.22 4.97
greedy selection [30] (N=99) 3.81 4.52 5.28
active learning [37] (N=99) 3.65 4.19 4.93
conventional AS [15] (N=99) 3.24 3.87 4.61
VAS (N=99) 4.61 5.64 6.55
We also present the results for large vehicle andship tar-
get class from DOTA dataset in the following Table 10 and
11 respectively. We see the proposed V AS performs notice-
ably better than all baselines, ranging from 16–56% relativeTable 9: ANT comparisons for the building target class.
Method C=12 C=15 C=18
random search (N=30) 5.54 7.18 8.58
greedy classification (N=30) 5.88 7.72 9.21
greedy selection [30] (N=30) 6.39 7.95 9.52
active learning [37] (N=30) 5.86 7.68 9.16
conventional AS [15] (N=30) 5.76 7.37 8.87
VAS (N=30) 7.56 9.02 10.41
random search (N=48) 4.97 6.41 7.66
greedy classification (N=48) 5.68 6.95 8.40
greedy selection [30] (N=48) 5.93 7.26 8.71
active learning [37] (N=48) 5.68 6.93 8.37
conventional AS [15] (N=48) 5.22 6.67 7.84
VAS (N=48) 6.85 8.29 9.65
random search (N=99) 4.35 5.37 6.44
greedy classification (N=99) 4.92 6.02 7.41
greedy selection [30] (N=99) 5.38 6.53 7.79
active learning [37] (N=99) 4.91 6.00 7.40
conventional AS [15] (N=99) 4.55 5.64 6.75
VAS (N=99) 6.75 8.27 9.46
to the state-of-the-art greedy selection approach. The exper-
imental outcomes in different settings are qualitatively sim-
ilar to the settings under Manhattan distance-based query
cost.
Table 10: ANT comparisons for the large vehicle target
class.
Method C=12 C=15 C=18
random search (N=36) 3.44 4.08 5.19
greedy classification (N=36) 3.95 4.62 5.56
greedy selection [30] (N=36) 4.18 4.86 5.89
active learning [37] (N=36) 3.92 4.60 5.54
conventional AS [15] (N=36) 3.71 4.22 5.28
VAS (N=36) 5.14 6.05 7.00
random search (N=64) 3.40 4.03 5.14
greedy classification (N=64) 3.87 4.59 5.55
greedy selection [30] (N=64) 3.99 4.77 5.67
active learning [37] (N=64) 3.85 4.54 5.51
conventional AS [15] (N=64) 3.61 4.12 5.26
VAS (N=64) 6.30 7.65 8.90
B. V AS Pseudocode
We have included the pseudocode of our proposed Visual
Active Search algorithm in table 1.
C. Policy architecture, training hyperparame-
ter, and the details of TTA
In table 12, we detail the VAS policy architecture with
number of target grids as N. We use a learning rate of 10−4,
11
Table 11: ANT comparisons for the ship target class.
Method C=12 C=15 C=18
random search (N=36) 2.69 3.38 4.46
greedy classification (N=36) 3.21 3.99 5.11
greedy selection [30] (N=36) 3.44 4.23 5.32
active learning [37] (N=36) 3.18 3.95 5.07
conventional AS [15] (N=36) 2.97 3.56 4.77
VAS (N=36) 4.58 5.34 6.23
random search (N=64) 2.54 3.01 4.21
greedy classification (N=64) 3.34 3.74 4.94
greedy selection [30] (N=64) 3.62 3.95 5.10
active learning [37] (N=64) 3.32 3.71 4.93
conventional AS [15] (N=64) 2.87 3.38 4.53