text
stringlengths
0
820
active learning [37] (N=48) 3.88 4.43 4.60
conventional active search [15] (N=48) 3.70 4.11 4.38
VAS (N=48) 5.61 9.26 12.15
random search (N=99) 1.55 2.99 4.18
greedy classification (N=99) 2.17 3.96 4.84
greedy selection [30] (N=99) 2.29 4.21 5.22
active learning [37] (N=99) 2.17 3.95 4.82
conventional active search [15] (N=99) 1.68 3.10 4.33
VAS (N=99) 4.29 6.91 8.98
The results are presented in Table 1 for the small car
class and in Table 2 for the building class. We see sub-
stantial improvements in performance of the proposed VAS
approach compared to all baselines, ranging from 15–260%
improvement relative to the most competitive state-of-the-
art approach, greedy selection . There are two general con-
sistent trends. First, as the number of grids Nincreases
compared to C(corresponding to sets of rows in either ta-
ble), performance of all methods declines, as the task be-
comes more challenging. However, the decline in perfor-mance is typically much greater for our baselines than for
VAS. Second, overall performance improves as Cincreases
(columns in both tables), and the relative advantage of VAS
increases, as it is better able to take advantage of the greater
budget than the baselines.
Figure 4: Comparison of policies learned using VAS(left) and the greedy
selection baseline method (right).
In Figure 5 we visually illustrate VAS search strategy in
comparison with the greedy selection baseline (the best per-
forming baseline). The plus signs correspond to success-
ful queries, to unsuccessful queries, and arrows repre-
sent query order. This shows that VAS quickly learns to
take advantage of the visual similarities between grids (af-
ter the first several failed queries, the rest are successful),
whereas our most competitive baseline— greedy selection —
fails to take advantage of such information. During the ini-
tial search phase, the V AS policy explores different types of
grids before exploiting grids it believes to have target ob-
jects.
Finally, we perform an ablation study to understand the
added value of including remaining budget Bas an input
in the VAS policy network. To this end, we modify the
combined feature representation of size (2N+1)×14×14,
consisting of input and auxiliary state features, each of size
N×14×14, and a single channel of size 14×14containing
the information of remaining search budget, as depicted in
Figure 3. We only eliminate the channel from the combined
feature representation that contains the information about
the number of queries left, resulting in 2N×14×14size
feature map. The resulting policy network is then trained
just as the original VAS architecture.
Table 3: Comparative ANT performance of VAS without remaining
search budget andVAS using small car as the target class.
Method C=25 C=50 C=75
VAS w/o remaining search budget (N=30) 4.47 7.38 9.62
VAS (N=30) 4.61 7.49 9.88
VAS w/o remaining search budget (N=48) 4.34 7.31 9.49
VAS (N=48) 4.56 7.45 9.63
VAS w/o remaining search budget (N=99) 2.63 4.29 5.69
VAS (N=99) 2.72 4.42 5.78
We compare the performance of the policy without re-
maining search budget (referred to as VAS without remain-
ing search budget ) with VAS in Table 3. Across all problem
sizes and search budgets, we observe a relatively small but
6
consistent improvement ( ∼1–3%) from using the remaining
search budget Bas an explicit input to the policy network.
5.2. Results on the DOTA Dataset
Next, we repeat 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 C∈{25,50,75}.
Table 4: ANT comparisons for the large vehicle target class on DOTA.
Method C=25 C=50 C=75
random search (N=36) 1.79 3.50 5.10
greedy classification (N=36) 2.64 4.07 5.88
greedy selection [30] (N=36) 2.82 4.21 5.97
active learning [37] (N=36) 2.63 4.06 5.84
conventional active search [15] (N=36) 1.92 3.63 5.34
VAS (N=36) 4.63 6.79 8.07
random search (N=64) 1.48 2.96 3.91
greedy classification (N=64) 2.59 3.77 5.48
greedy selection [30] (N=64) 2.72 4.10 5.77
active learning [37] (N=64) 2.57 3.74 5.47
conventional active search [15] (N=64) 1.64 3.15 4.23
VAS (N=64) 5.33 8.47 10.51
Table 5: ANT comparisons for the shiptarget class on the DOTA dataset.
Method C=25 C=50 C=75
random search (N=36) 1.73 3.07 4.26
greedy classification (N=36) 2.04 3.65 4.92
greedy selection [30] (N=36) 2.33 3.84 5.01
active learning [37] (N=36) 2.01 3.64 4.91
conventional active search [15] (N=36) 1.86 3.25 4.40
VAS (N=36) 3.31 5.34 6.74
random search (N=64) 1.26 2.33 3.14
greedy classification (N=64) 1.89 3.06 3.75
greedy selection [30] (N=64) 2.07 3.32 4.02
active learning [37] (N=64) 1.87 3.05 3.72
conventional active search [15] (N=64) 1.41 2.48 3.38
VAS (N=64) 3.58 6.38 7.83
The results are presented in Tables 4 and 5, and are