gremlin97/RemoteSensingCorpus · Datasets at Hugging Face

text stringlengths 0 820
MoCo-V2+Geo+TP ResNet50 69.56 (+6.43) 72.76 (+6.43)
Table 2: Experiments on fMoW on classifying temporal
data. In the table, we compare the results to the ones on
single image classiﬁcation. Here we present results corre-
sponding to linear classiﬁcation on frozen features only.
5.2. Transfer Learning Experiments
Previously, we performed pre-training experiments on
fMoW dataset and quantiﬁed the quality of the representa-
tions by supervised training a linear layer for image recogni-
tion on fMoW. In this section, we perform transfer learning
experiments on different low level tasks.
5.2.1 Object Detection
For object detection, we use the xView dataset [16] consist-
ing of high resolution satellite images captured with similar
sensors to the ones in the fMoW dataset. The xView dataset
7
pre-train AP50↑
Random Init. 10.75
Sup. Learning (IN wts. init.) 14.44
Sup. Learning (Scratch) 14.42
MoCo-V2 15.45 (+4.70)
MoCo-V2-Geo 15.63 (+4.88)
MoCo-V2-TP 17.65 (+6.90)
MoCo-V2-Geo+TP 17.74 (+6.99)
Table 3: Object detection results on the xView dataset.
consists of 846 very large ( ∼2000×2000 pixels) satellite
images with bounding box annotations for 60 different class
categories including airplane, passenger vehicle, maritime
vessel, helicopter etc.
Implementation Details We ﬁrst divide the set of large
images into 700 training and 146 test images. Then, we
process the large images to create 416 ×416 pixels images
by randomly sampling the bounding box coordinates of the
small image and we repeat this process 100 times for each
large image. In this process, we ensure that there is less than
25% overlap between any two bounding boxes from the
same image. We then use RetinaNet [18] with pre-trained
ResNet-50 backbone and ﬁne-tune the full network on the
xView training set. To train RetinaNet, we use learning rate
of 1e-5 and a batch size of 4 and Adam optimizer.
Qualitative Analysis Table 3 shows the object detection
performance on the xView test set. We achieve the best re-
sults with the addition of temporal positive pair, and geo-
location classiﬁcation pre-text task into MoCo-v2. With
our ﬁnal model, we can outperform the randomly initial-
ized weights by 7%AP and the supervised learning on the
fMoW by 3.3%AP.
5.2.2 Image Segmentation
In this section, we perform downstream experiments on the
task of Semantic Segmentation on SpaceNet dataset [40].
The SpaceNet datasets consists of 5000 high resolution
satellite images with segmentation masks for buildings.
Implementation Details We divide our SpaceNet dataset
into training and test sets of 4000 and 1000 images respec-
tively. We use PSAnet [50] network with ResNet-50 back-
bone to perform semantic segmentation. We train PSAnet
network with a batch size of 16 and a learning rate of 0.01
for 100 epochs and use SGD optimizer.
Qualitative Analysis Table 4 shows the segmentation per-
formance of differently initialized backbone weights on the
SpaceNet test set. Similar to object detection, we achieve
the best IoU scores with the addition of temporal positives
and geo-location classiﬁcation task. Our ﬁnal model out-
performs the randomly initialized weights and supervised
learning by 3.58% and2.94% IoU scores. We observe that
the gap between the best and worst models shrinks goingfrom the image recognition to object detection, and seman-
tic segmentation task. This aligns with the performance of
the MoCo-v2 pre-trained on ImageNet and ﬁne-tuned on
the Pascal-VOC object detection and semantic segmenta-
tion experiments [13, 3].
pre-train mIOU↑
Random Init. 74.93
Imagenet Init. 75.23
Sup. Learning (IN wts. init.) 75.61
Sup. Learning (Scratch) 75.57
MoCo-V2 78.05 (+3.12)
MoCo-V2-Geo 78.42 (+3.49)
MoCo-V2-TP 78.48 (+3.55)
MoCo-V2-Geo+TP 78.51 (+3.58)
Table 4: Semantic segmentation results on Space-Net.
pre-train Top-1 Accuracy ↑
Random Init. 51.89
Imagenet Init. 53.46
Sup. Learning (IN wts. init.) 54.67
Sup. Learning (Scratch) 54.46
MoCo-V2 55.18 (+3.29)
MoCo-V2-Geo 58.23 (+6.34)
MoCo-V2-TP 57.10 (+5.21)
MoCo-V2-Geo+TP 57.63 (+5.74)
Table 5: Land Cover Classiﬁcation on NAIP dataset.
5.2.3 Land Cover Classiﬁcation
Finally, we perform transfer learning experiments on land
cover classiﬁcation across 66 land cover classes using high
resolution remote sensing images obtained by the USDA’s
National Agricultural Imagery Program (NAIP). We use the
images from the California’s Central Valley for the year of
2016. Our ﬁnal dataset consists of 100,000 training and
50,000 test images. Table 5 shows that our method outper-
forms the randomly initialized weights by 6.34% and super-
vised learning by 3.77%.

MoCo-V2+Geo+TP ResNet50 69.56 (+6.43) 72.76 (+6.43)

Table 2: Experiments on fMoW on classifying temporal

data. In the table, we compare the results to the ones on

single image classiﬁcation. Here we present results corre-

sponding to linear classiﬁcation on frozen features only.

5.2. Transfer Learning Experiments

Previously, we performed pre-training experiments on

fMoW dataset and quantiﬁed the quality of the representa-

tions by supervised training a linear layer for image recogni-

tion on fMoW. In this section, we perform transfer learning

experiments on different low level tasks.

5.2.1 Object Detection

For object detection, we use the xView dataset [16] consist-

ing of high resolution satellite images captured with similar

sensors to the ones in the fMoW dataset. The xView dataset

7

pre-train AP50↑

Random Init. 10.75

Sup. Learning (IN wts. init.) 14.44

Sup. Learning (Scratch) 14.42

MoCo-V2 15.45 (+4.70)

MoCo-V2-Geo 15.63 (+4.88)

MoCo-V2-TP 17.65 (+6.90)

MoCo-V2-Geo+TP 17.74 (+6.99)

Table 3: Object detection results on the xView dataset.

consists of 846 very large ( ∼2000×2000 pixels) satellite

images with bounding box annotations for 60 different class

categories including airplane, passenger vehicle, maritime

vessel, helicopter etc.

Implementation Details We ﬁrst divide the set of large

images into 700 training and 146 test images. Then, we

process the large images to create 416 ×416 pixels images

by randomly sampling the bounding box coordinates of the

small image and we repeat this process 100 times for each

large image. In this process, we ensure that there is less than

25% overlap between any two bounding boxes from the

same image. We then use RetinaNet [18] with pre-trained

ResNet-50 backbone and ﬁne-tune the full network on the

xView training set. To train RetinaNet, we use learning rate

of 1e-5 and a batch size of 4 and Adam optimizer.

Qualitative Analysis Table 3 shows the object detection

performance on the xView test set. We achieve the best re-

sults with the addition of temporal positive pair, and geo-

location classiﬁcation pre-text task into MoCo-v2. With

our ﬁnal model, we can outperform the randomly initial-

ized weights by 7%AP and the supervised learning on the

fMoW by 3.3%AP.

5.2.2 Image Segmentation

In this section, we perform downstream experiments on the

task of Semantic Segmentation on SpaceNet dataset [40].

The SpaceNet datasets consists of 5000 high resolution

satellite images with segmentation masks for buildings.

Implementation Details We divide our SpaceNet dataset

into training and test sets of 4000 and 1000 images respec-

tively. We use PSAnet [50] network with ResNet-50 back-

bone to perform semantic segmentation. We train PSAnet

network with a batch size of 16 and a learning rate of 0.01

for 100 epochs and use SGD optimizer.

Qualitative Analysis Table 4 shows the segmentation per-

formance of differently initialized backbone weights on the

SpaceNet test set. Similar to object detection, we achieve

the best IoU scores with the addition of temporal positives

and geo-location classiﬁcation task. Our ﬁnal model out-

performs the randomly initialized weights and supervised

learning by 3.58% and2.94% IoU scores. We observe that

the gap between the best and worst models shrinks goingfrom the image recognition to object detection, and seman-

tic segmentation task. This aligns with the performance of

the MoCo-v2 pre-trained on ImageNet and ﬁne-tuned on

the Pascal-VOC object detection and semantic segmenta-

tion experiments [13, 3].

pre-train mIOU↑

Random Init. 74.93

Imagenet Init. 75.23

Sup. Learning (IN wts. init.) 75.61

Sup. Learning (Scratch) 75.57

MoCo-V2 78.05 (+3.12)

MoCo-V2-Geo 78.42 (+3.49)

MoCo-V2-TP 78.48 (+3.55)

MoCo-V2-Geo+TP 78.51 (+3.58)

Table 4: Semantic segmentation results on Space-Net.

pre-train Top-1 Accuracy ↑

Random Init. 51.89

Imagenet Init. 53.46

Sup. Learning (IN wts. init.) 54.67

Sup. Learning (Scratch) 54.46

MoCo-V2 55.18 (+3.29)

MoCo-V2-Geo 58.23 (+6.34)

MoCo-V2-TP 57.10 (+5.21)

MoCo-V2-Geo+TP 57.63 (+5.74)

Table 5: Land Cover Classiﬁcation on NAIP dataset.

5.2.3 Land Cover Classiﬁcation

Finally, we perform transfer learning experiments on land

cover classiﬁcation across 66 land cover classes using high

resolution remote sensing images obtained by the USDA’s

National Agricultural Imagery Program (NAIP). We use the

images from the California’s Central Valley for the year of

2016. Our ﬁnal dataset consists of 100,000 training and

50,000 test images. Table 5 shows that our method outper-

forms the randomly initialized weights by 6.34% and super-

vised learning by 3.77%.