Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
4%) in search performance by leveraging TTA in our pro-
posed V AS framework. Specifically, the performance gap
becomes more noticeable as the search budget increases.
We observe a similar trend when we train V AS with building
as target and evaluate using small car as target as presented
in table 23. Such results reinforce the importance of TTA in
scenarios (especially when the search budget is large) when
the search target differs between training and execution en-
vironments.
Table 22: Comparative results on xView dataset with small car
andBuilding as the target class during training and inference re-
spectively under uniform query cost setting.
Method C=18 C=24 C=30 C=60
without TTA (N=900) 3.32 4.30 5.41 10.39
Stepwise TTA (N=900) 3.38 4.37 5.54 10.68
Online TTA (N=900) 3.41 4.42 5.60 10.81
Table 23: Comparative results on xView dataset with building
andsmall car as the target class during training and inference re-
spectively under uniform query cost setting.
Method C=18 C=24 C=30 C=60
without TTA (N=900) 1.61 2.07 2.60 4.93
Stepwise TTA (N=900) 1.63 2.10 2.66 5.04
Online TTA (N=900) 1.66 2.15 2.71 5.12
15
(a) The original image
step 1
step 3
step 5
step 7
step 9
step 11
step 13
step 15
(b)(Top row )Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS while enforcing
the query outcomes at every stage being “ unsuccessful ”.(Bottom row )Query sequences, and corresponding heat maps (darker indicates
higher probability), obtained using V AS while enforcing the query outcomes at every stage being “ successful ”.
Figure 10: Sensitivity Analysis of VAS with a sample test image and large vehicle as target class under uniform query cost.
J. Saliency map visualization of VAS
In Figure (16,17,18), we show the saliency maps ob-
tained using a pre-trained V AS policy at different stages of
the search process. Note that, at every step, we obtain the
saliency map by computing the gradient of the output that
corresponds to the query index with respect to the input.
Figure 16 corresponds to the large vehicle target class while
the Figure 17 and Figure 18 correspond to the small vehi-
cle. All saliency maps were obtained using the same searchbudget (K = 15). These visualizations capture different as-
pects of the V AS policy. Figure 16 shows its adaptability,
as we see how heat transfers from non-target grids to the
grids containing targets as search progresses. By compar-
ing saliency maps at different stages of the search process,
we see that, V AS explores different regions of the image at
different stages of search, illustrating that our approach im-
plicitly trades off exploration and exploitation in different
ways as search progresses. Figure 17 shows the effect of su-
pervised training on V AS policy. If we observe the saliency
16
(a) The original image
step 1
step 3
step 5
step 7
step 9
step 11
step 13
step 15
(b)(Top row )Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS while enforcing
the query outcomes at every stage being “ unsuccessful ”.(Bottom row )Query sequences, and corresponding heat maps (darker indicates
higher probability), obtained using V AS while enforcing the query outcomes at every stage being “ successful ”.
Figure 11: Sensitivity Analysis of VAS with a sample test image and caras target class under uniform query cost.
maps across time, we see that V AS never searches for small
vehicles in the sea, having learned not to do this from train-
ing with similar images. Additionally, we notice that the
saliency map’s heat expands from left to right as the time
step increases, encompassing more target grids, leading to
the discovery of more target objects. We observe similar
phenomena in figure 18. We can see that while earlier in
the search process queries tend to be less successful, as the
search evolves, our approach successfully identifies a clus-ter of grids that contain the desired object, exploiting spatial
correlation among them. Additionally, at different stages of
the search process, V AS identifies different clusters of grids
that include the target object.
17
(a) The original image
step 1
step 3
step 5
step 7
step 9
step 11
step 13
step 15
(b)(Top row )Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS while enforcing
the query outcomes at every stage being “ unsuccessful ”.(Bottom row )Query sequences, and corresponding heat maps (darker indicates
higher probability), obtained using V AS while enforcing the query outcomes at every stage being “ successful ”.
Figure 12: Sensitivity Analysis of VAS with a sample test image and ship as target class under uniform query cost.
18
(a) The original image
step 1
step 3
step 5