text
stringlengths
0
820
broadly consistent with our observations on the xView
dataset, with VAS outperforming all baselines by ∼40−
80%, with the greatest improvement typically coming with
a higher search budget C.
5.3. Visualization of V AS Strategy
In figure 5, we demonstrate the sequential behavior of a
pretrained V AS policy during inference. We have shaded
the grid such that darker indicates higher probability. The
darkest grid at each step is the target to be revealed. In the
first row we have V AS searching for large vehicles . In step
1, V AS looks at the roof of a building, which looks very
similar to a large vehicle in the overhead view. Next in step
3, it searches a grid with a large air conditioner which also
looks similar to a large vehicle . Having viewed these twoconfusers, V AS now learns that the rest of the grids with
building roofs likely contain no large vehicles . It is impor-
tant to note that this eliminates a large portion of the middle
of the image from consideration as it is entirely roof tops.
In step 5, it moves to an area which is completely different,
a road where it finds a large vehicle . V AS now aggressively
begins examining grids with roads. In steps 7 through 13
it searches roads discovering large vehicle . Finally in step
15 it explores to a parking lot containing a large vehicle . In
our middle example we have V AS targeting small cars . In
step 1, V AS targets a road and fails to find a car. In step
3, it searches another road in a different region and finds a
car. Having explored regions with prominent major roads
it moves to a parking lot in step 5 and finds a car. It now
searches a similar parking lot in step 7. Having explored
grids with parking lots it goes back to searching minor roads
for the duration of its search. V AS does not visit a parking
lot in the north east corner, but this parking lot is visually
much different from the other two (i.e. it’s not rectangular).
In our bottom example we have V AS searching for ships .
In step 1, V AS searches near a harbor. Having found a ship
it begins exploring similar harbor regions. In step 3 and
5 it searches other parts of the same harbor finding ships .
In steps 7-9, it searches areas similar to the harbor with-
outships . V AS now learns that ships are not likely present
in the rest of the dock and explores different regions leav-
ing the rest of the dock unexplored. These three examples
demonstrate V AS’s tendency for the explore-exploit behav-
ior typical of reinforcement learning algorithms. Addition-
ally, we note that V AS has an ability to eliminate large areas
that would otherwise confuse standard greedy approaches.
5.4. Efficacy of Test-Time Adaptation
One of the important features of the visual active search
problem is that queries actually allow us to observe partial
information about target labels at inference time . Here, we
evaluate how our approaches to TTA that take advantage
of this information perform compared to the baseline VAS
without TTA , as well as state-of-the-art TTA baselines dis-
cussed in Section 4.2 (where FixMatch is adapted to also
take advantage of observed labels).
Consider first the case where there is no difference be-
tween training and test distribution over classes. As before
we consider xView and DOTA for analysis. The results are
presented in Figure 6, and show a consistent pattern. The
TTT approach performs the worst, followed by (out adap-
tation of) FixMatch , which is only slightly better than TTT.
Stepwise TTA outperforms both TTT andFixMatch , albeit
slightly, and Online TTA is, somewhat surprisingly much
better than all others (this is surprising since it has a lower
frequency of model update compared to Stepwise TTA ).
Finally, we consider a TTA setting in which the domain
exhibits a non-trivial distributional shift at inference time.
7
step 1
step 3
step 5
step 7
step 9
step 11
step 13
step 15
Figure 5: Query sequences, and corresponding heat maps (darker indicates higher probability), obtained using V AS for different target types.
Figure 6: Comparative results of TTA methods on V AS framework.
xView (top; small car target): (left) N=48, (right) N=99. DOTA
(bottom; large vehicle target): (left) N=36, (right) N=64.
In this case, we would expect the conventional TTT andFix-
Match methods to be more competitive, as they have been
specifically designed to account for distribution shift. We
model distribution shift by training the search policy using
one target object, and then applying it in the decision con-
text for another target object. Specifically, for xView, we
usesmall car as the target class during training, and build-
ingas the target class at test time. Similarly, on the DOTA
dataset we use large vehicle as the target class at training
time, and use ship as the target at test time.
The results for the TTA setting with distribution shift are
presented in Table 6 and 7 for the xView and the DOTA
dataset respectively, where we also add a comparison to the
VAS without TTA of any kind. We observe that the results
here remain consistent, with the proposed Online TTA out-
performing the other approaches, with Stepwise TTA yield-
ing the second-best performance.Table 6: Comparative results on xView dataset with small car andBuild-
ingas the target class during training and inference respectively.
Method C=25 C=50 C=75
without TTA (N=30) 5.28 8.58 11.42
TTT [27] (N=30) 5.30 8.61 11.45
FixMatch [26] (N=30) 5.31 8.62 11.47