Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
VAS (N=64) 5.04 6.50 7.38
Algorithm 1 The VAS algorithm.
Require: A search task instance (xi, yi); budget constraint C;
search policy ψ(xi, o, B)with parameters θ;
1:Initialize o0=[0...0];B0=C; step t=0
2:while Bt>0do
3: ˜y=ψ(xi, ot, Bt)
4: j← /leftr⫯g⊸tl⫯neSamplej∈{Unexplored Grids }[˜y]
5: Query grid cell with index jand observe true label y(j).
6: Obtain reward Rt=y(j).
7: Update ottoot+1witho(j)=2y(j)−1.
8: Update BttoBt+1withBt+1=Bt−c(k, j)(assuming
we query k’th grid at (t−1)).
9: Collect transition tuple ( τ) at step t, i.e., τt=(state
=(xi, ot, Bt), action = j, reward = Rt, next state =
(xi, ot+1, Bt+1) ).
10: t← /leftr⫯g⊸tl⫯net+1
11:end while
12:Update the search policy parameters, i.e., θusing REIN-
FORCE objective as in 3 based on the collected transition
tuples ( τt) throughout the episode.
13:Return updated search policy parameters, i.e., θ.
batch size of 16, number of training epochs 200, and the
Adam optimizer to train the policy network in all results.
We add a self-supervised head rto the V AS policy archi-
tecture for TTT. The architecture of self-supervised head is
detailed in table 13. We applied a series of 4 up-convolution
layers with intermediate ReLU activations followed by a
tanh activation layer on the semantic features extracted us-
ing ResNet34. For FixMatch, our V AS architecture remains
unchanged, and we apply only spatially invariant augmen-
tations (e.g auto contrast, brightness, color, and contrast)
and ignore all translation augmentations (translate X, trans-
late Y , ShearX etc.) to obtain the augmented version of the
input image. We update the model parameters after every
query step using a cross-entropy loss between a pseudo-
target and a predicted vector as described below. We define
the pseudo-target vector as follows. Whenever a query jissuccessful ( yj=1), we construct a label vector as the one-
hot vector with a 1 in the jth grid cell. However if yj=0, we
associate each queried grid cell with a 0, and assign a uni-
form probability distribution over all unqueried grids. Pre-
diction vector is the “logit” representation obtained from the
V AS policy. We used the Adam optimizer with a learning
rate of 10−4for both TTT and FixMatch.
Table 12: V AS Policy Architecture
Layers Configuration o/p Feature Map size
Input RGB Image 3 ×2500×3000
Feat. Extraction ResNet-34 512 ×14×14
Conv1 c:N k: 1×1 N×14×14
Tile1 Grid State ( o) N×14×14
Tile2 Query Left ( B) 1×14×14
Channel Concat Conv1,Tile1,Tile2 (2N+1)×14×14
Conv2 c:3 k: 1×1 3×14×14
Flattened Conv2 588
FC1+ReLU ( 588−>2N) 2N
FC2 ( 2N−>N) N
Table 13: Self-supervised head Architecture
Layers Configuration
Input: Latent Feature 36×14×14
1st Up-conv layer in-channel:36;out-channel:36;k: 3×3;stride:2;padd:0
Activation Layer ReLU
2nd Up-conv layer in-channel:36;out-channel:24;k: 3×3;stride:2;padd:1
Activation Layer ReLU
3rd Up-conv layer in-channel:24;out-channel:12;k: 2×2;stride:4;padd:1
Activation Layer ReLU
4th Up-conv layer in-channel:12; out-channel:3; k: 2×2; stride:2; padd:0
Normalization layer tanh
D. Search Performance Comparison with Dif-
ferent Feature Extractor Module
In this section, we compare the performance of V AS with
different feature extraction module. We use state-of-the-art
feature extraction modules, such as ViT [6] and DINO [4]
for comparison. The Vision Transformer (ViT) [6] is a
transformer encoder model (BERT-like) pretrained on a
large collection of images in a self-supervised fashion,
namely ImageNet-21k (a collection of 14 million images),
at a resolution of 224×224pixels, with patch resolution
of16×16. Note that, we use off the shelf pretrained ViT
model provided by huggingface (google/vit-base-patch16-
224-in21k). We call the resulting policy VAS-ViT . Similar to
12
ViT, DINO [4] is also based on transformer encoder model.
Images are presented to the DINO model as a sequence of
fixed-size patches (resolution 8x8), which are linearly em-
bedded. For our experiment, we use DINO pretrained on
ImageNet-1k, at a resolution of 224x224 pixels. For our ex-
periments, we use pretrained DINO model provided by hug-
gingface (facebook/dino-vits8). We call the resulting policy
as V AS-DINO. In table 14, 15 we report the performance
of V AS-ViT and V AS-DINO and compare them with V AS.
Table 14: ANT comparisons with different feature extraction
module for the small car target class on xView.
Method C=25 C=50 C=75
VAS-DINO (N=30) 4.56 7.41 9.83
VAS-ViT (N=30) 4.64 7.47 9.86
VAS (N=30) 4.61 7.49 9.88
VAS-DINO (N=48) 4.52 7.41 9.59
VAS-ViT (N=48) 4.56 7.44 9.68
VAS (N=48) 4.56 7.45 9.63
Table 15: ANT comparisons with different feature extraction