text
stringlengths
0
820
UperNet SeCo-ResNet-50 57.2 63.9 71.7 66.9 69.9 54.5 45.9 38.9 58.2 44.8 33.2 9.3 52.3 71.6 83.3 51.4
UperNet RSP-ResNet-50 61.6 64.2 75.9 68.8 69.9 58.5 54.4 40.2 59.6 47.5 32.1 43.8 65.4 76.5 82.8 51.5
UperNet IMP-Swin-T 64.6 69.2 76.5 74.1 69.9 56.3 60.1 41.9 62.3 51.6 44.7* 45.8 64.5 75.9 85.7 56.7
UperNet RSP-Swin-T 64.1 67.0 74.6 73.7 70.7 59.0 60.1 44.3 62.0 50.6 37.6 46.8 64.9 76.2 85.2 53.8
UperNet IMP-ViTAEv2-S 65.3* 71.4* 77.5* 68.2 71.0 60.8 61.9 43.0 63.8* 53.6* 43.4 44.8 65.1 77.9* 86.4* 57.7*
UperNet RSP-ViTAEv2-S 64.3 71.3 74.3 72.2 70.4 57.4 63.0* 44.0 62.5 51.6 35.4 47.0 62.2 77.7 85.2 54.7
1ST: storage tank. BD: baseball diamond. TC: tennis court. BC: baseball court. GTF: ground track field. LV: large vehicle. SV: small vehicle. HC: helicopter.
SP: swimming pool. RA: roundabout. SBF: soccer ball field.
high-resolution features have not encoded sufficient high-level
semantics, while LANet [92] not only simultaneously enhance
the high and low-level features, it also enriches the semantics
of the high-resolution features. Thus, the segmentation perfor-
mance of the evaluated models based on UperNet on small
objects, such as cars, needs to be improved. On the other
hand, the IMP-Swin-T performs to be competitive and the
IMP-ViTAEv2-S achieves the best performance on the iSAID
dataset, outperforming the SOTA methods such as the HRNet
[103] and OCR [106] as well as a series of methods that
specially designed for aerial semantic segmentation, e.g., the
FarSeg [108] and FactSeg [109].
Table VII also shows the advantages of RSP models lying
in the “Bridge” category, which conforms to the finding in
the previous scene recognition task. Nevertheless, we can also
see from Table VI-VII that, on the segmentation task, theperformances of RSP are not as good as the classical IMP.
In our considerations, there may be two reasons. The first
one is the difference between the pretraining dataset and the
evaluation one. Besides the dataset volume (note that the train-
ing sample and category numbers of MillionAID are smaller
than ImageNet-1k), the spectral disparities also have a side
impact on the performance, especially on the Potsdam dataset,
which adopts the IR-R-G channels instead of the ordinary
RGB image (See Figure 6). Another reason we attribute to
the difference between tasks. The representation used for scene
recognition needs to have a global understanding of the whole
scene as Figure 4 shows, while the segmentation task requires
the features to be more detailed while possessing high-level
semantic information simultaneously since they separately
conduct the scene-level or pixel-level classification. To prove
this conjecture, we then evaluate these networks on the aerial
12 JOURNAL OF L ATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
(a) (b) (c) (d)
(e) (f) (g) (h)Imper. surf.
Building
Low veg.
Tree
Car
Ignore
Fig. 7. Segmentation maps of the UperNet with different backbones on the Potsdam dataset. (a) Ground Truth. (b) IMP-ResNet-50. (c) SeCo-ResNet-50. (d)
RSP-ResNet-50. (e) IMP-Swin-T. (f) RSP-Swin-T. (g) IMP-ViTAEv2-S. (h) RSP-ViTAEv2-S.
object detection task in the next section. The granularity of
the representation needed for detection probably lies between
those for the segmentation and recognition tasks, since one of
the aims in the detection task is the object-level classification.
Qualitative Results and Analyses: We present some visual
segmentation results of the UperNet with different backbones
on the Potsdam dataset in Figure 7. As can be seen, only
the ViTAEv2-S successfully connects the long strip low veg-
etations (see the red boxes), while IMP-ViTAEv2-S performs
slightly better than RSP-ViTAEv2-S, which is consistent with
the quantitative results in Table VI.
C. Aerial Object Detection
Since the aerial images are top-down photoed in the sky,
the objects can be presented in any direction in the birdview.
Thus, the aerial object detection is the oriented bounding
box (OBB) detection, which is distinguished from the usual
horizontal bounding box (HBB) task on natural images [111],
[117], [134]. In this paper, similar to segmentation, we also
use different detection datasets in the experiments. Concretely,
we evaluated on the multi-category RS objects detection and
the single-category ship detection subtasks, respectively.
1) Dataset: Two datasets including the large-scale DOTA
[135] scenes and the commonly used HRSC2016 [136] dataset
are separately utilized for the above objectives.
•DOTA: This is the most famous large-scale dataset for
OBB detection. It totally contains 2,806 images whose
size ranges from 800 ×800 to 4,000 ×4,000, where
188,282 instances belonging to 15 categories are in-
cluded. The training, validation, and testing set separately
have 1,411/458/937 tiles. It should be noticed that the cat-
egories are completely the same with the iSAID dataset,since the two datasets share the same set of scenes. The
difference lies in the annotations for different tasks.
•HRSC2016: This is a specialized ship detection dataset,
where the bounding boxes are annotated in arbitrary ori-
entations. 1,061 images with the size ranging from 300 ×
300 to 1,500 ×900 are included. In the official division,
436/181/444 images are used for training, validation, and
testing, respectively. The dataset only has one category,
since there is no need to recognize the type of ships.
2) Implementation Detail and Experimental Setting: Simi-
lar to segmentation, the ResNet models are trained using the
SGDM algorithm with a learning rate of 0.005, a momentum
of 0.9, and a weight decay of 0.0001, while the vision
transformers are trained with the AdamW optimizer, where the
learning rate and weight decay are separately set to 0.0001 and
0.05. These models are trained for 12 and 36 epochs with a
batch size of 2 on DOTA and HRSC2016 scenes, respectively.
The learning rate is adjusted by a multi-step scheduler. On the
DOTA dataset, the learning rate will be separately reduced by
10×after the 8th epoch and the 11th epoch, while on the
HRSC2016 scene, the corresponded settings are epoch 24 and