Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
In addition, we observe that compared with IMP-ViTAEv2-
S, the APs of most categories obtained by RSP-ViTAEv2-S
are smaller, implying the universality of IMP. Nevertheless,
the mAP of RSP-ViTAEv2-S is higher than IMP-ViTAEv2-
S, since RSP has significant advantages in the categories
of “Bridge” and aerial vehicles including “Helicopter” or
“Plane”, echoing the previous finding in the segmentation
experiments. While on the other categories, the gaps between
these two models are not very large. Combining the above two
points, RSP-ViTAEv2-S delivers better overall performance
than IMP-ViTAEv2-S. On HRSC2016 dataset, the CHPDet
[133] that performs the best is a specifically designed detector
by considering the ship characteristics. For ORCN [126]
related networks, the results of RSP and IMP are roughly
the same, where there are wins or losses on both sides.
Compared with CNN, the vision transformer models have not
demonstrated the advantages. We think that on this relatively
easy subtask, where only one category needed to be detected
and the ship sizes in HRSC2016 are relatively larger than
DOTA, the performance is probably saturated.
Qualitative Results and Analyses: We visualize some
detection results of the ORCN model with the ViTAEv2-S
backbones on the DOTA testing set in Figure 8. The red boxes
show that, when objects are densely distributed, the RSP-
14 JOURNAL OF L ATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
IMP
RSPLV & SV BR
Fig. 8. Visual detection results of the ORCN model with the ViTAEv2-S
backbones on the DOTA testing set. LV: large vehicle. SV: small vehicle.
BR: Bridge. IMP: IMP-ViTAEv2-S. RSP: RSP-ViTAEv2-S.
ViTAEv2-S can still predict correct object categories, while
the IMP-ViTAEv2-S is confused by the dense context and
makes wrong predictions. For the “Bridge” category, the IMP-
ViTAEv2-S produces missing detections (see yellow boxes),
while the RSP-ViTAEv2-S model successfully detects the long
bridge with a high confidence score.
D. Aerial Change Detection
We then apply the above models on a typical application
in the RS field, i.e., change detection, which aims to find the
changes between two aerial images of a same region captured
at different times. It is formulated as a pixel-level binary
classification task, where “1” indicates change.
1) Dataset: We adopt the commonly used CDD [137] and
LEVIR [138] datasets to comprehensively evaluate the above
models on this task since they separately involve the natural
and artificial changes.
•CDD: The original dataset contains 11 pairs of multi-
source real season-varying RS images collecting from
GE, where 7 pairs of images are at the size of 4,725
×2,200 and 4 image pairs are at the size of 1,900 ×
1,000 pixels. The resolutions are ranging from 0.03m
to 1m. Then, [139] clipped the images to a series of
256×256 patches and generated a dataset, where
the sizes of the training, validation and testing set are
10,000/3,000/3,000, respectively.
•LEVIR: This dataset is collected using the GE API on 20
different regions in Texas, the USA, from 2002 to 2018. It
contains 637 image pairs at the size of 1,024 ×1,024 and
with a high resolution of 0.5m, where most changes are
from man-made structures, including 31,333 independent
building change entities. The training, validation, and
testing set contain 445/64/128 image pairs, respectively.TABLE X
RESULTS OF THE BIT CHANGE DETECTION MODEL WITH DIFFERENT
BACKBONES AND SOTA METHODS ON THE TESTING SET OF CDD AND
LEVIR DATASETS .†: THE RESULT IS FROM THE ORIGINAL BIT PAPER .
Method BackboneF1 score
CDD LEVIR
FC-EF [141] — 77.11 62.32
FC-Siam-conc [141] — 82.50 68.21
FC-Siam-diff [141] — 83.73 63.09
CLNet [142] — 92.10 90.00
SNUNet-c48 [143] — 96.20 —
IFN [144] IMP-VGG-16 90.30 83.57
DASNet [145] IMP-VGG-16 91.93 82.83
SRCDNet [146] IMP-ResNet-18 90.02 —
STANet [138] IMP-ResNet-18 90.75 87.34
BSFNet [147] IMP-ResNet-18 91.90 88.00
DSAMNet [148] IMP-ResNet-18 93.69 —
HRTNet [149] IMP-HRNet-W18 93.71 88.48
CDNet+IAug [150] IMP-ResNet-18 — 89.00
BIT†[140] IMP-ResNet-18 — 89.31
ChangeFormer [151] IMP-MiT-B2 [47] — 90.40
CS-HSNet [152] IMP-ResNet-50 94.95 90.79
LSS-Net [153] IMP-SE-ResNet-50 96.30 —
ChangeStar [154] IMP-ResNext-101-32 ×4d — 91.25
BIT IMP-ResNet-50 95.09 89.19
BIT SeCo-ResNet-50 95.95 90.14
BIT RSP-ResNet-50 96.00 90.10
BIT IMP-Swin-T 94.77 90.25
BIT RSP-Swin-T 95.21 90.10
BIT IMP-ViTAEv2-S 97.02* 91.26*
BIT RSP-ViTAEv2-S 96.81 90.93
2) Implementation Detail and Experimental Setting: In this
section, we adopt a SOTA framework — BIT [140], which
uses the transformer to capture the contextual information
between different temporal images for change detection. If
BIT is equipped with the ResNet backbone, it is optimized
by the SGDM optimizer, where the learning rate, momentum,
and weight decay are 0.001, 0.99, and 0.0005. While the Swin
or ViTAE based BIT models are trained with the AdamW