text
stringlengths
0
820
(a) (b) (c)Fig. 6. Visual samples from different datasets: (a) MillionAID, (b) Potsdam,
and (c) iSAID.
2) Implementation Detail and Experimental Setting: We
adopt different settings for the above backbone networks by
following the common practice. Concretely, the ResNet based
networks are trained with the mini-batch stochastic gradient
descent with momentum (SGDM) strategy, where the initial
learning rate, weight decay, and momentum are separately set
to 0.01, 0.0005, and 0.9. The learning rate is optimized by the
polynomial scheduler: currentlr=minlr+initiallr·(1−
iter
max iter)power, whereminlr= 0.0001,power = 0.9. While
the vision transformers, such as the Swin-T and ViTAEv2-
S, are trained using the AdamW optimizer, whose learning
rate and weight decay are 6e-5 and 0.01. The learning rate
schedule adopts the polynomial decay policy with a power
of 1.0 and minlrof 0. They also have a linear warming
up stage at the first 1,500 iterations with the initial learning
rate of 1e-6. For a fair comparison, all networks are training
80k iterations with a batch size of 8. Following [13], [14], we
use the UperNet [15] as the unified segmentation framework
for all the pretrained backbones since the output stride equals
32. All methods are implemented based on mmsegmentation
[82]. For the convenience of training, the Potsdam and iSAID
are separately sampled and cropped into patches with a size
of 512 ×512 and 896 ×896 with a stride of 384 and
512, respectively. We use the random horizontal flipping data
augmentation strategy. Following the evaluation protocol in the
aerial segmentation community, for the Potsdam dataset, we
report the OA, mean F1 score (mF1), and per class F1 score.
Note that the clutter category is regarded as the background
and ignored during computing loss and evaluation metrics.
While for the iSAID dataset, the intersection over union (IoU)
of foreground categories and the average IOU of all classes
(including background) are calculated. All evaluations are
conducted on a single scale for a fair comparison.
3) Experimental Results: Quantitative Results and Anal-
yses: Table VI-VII present the segmentation results of our
methods and other SOTA methods. It can be seen that when
changing the backbone from ResNet-50 to Swin-T, and then
to ViTAEv2-S, the performance is increased. The results are
consistent with the aforementioned scene recognition results,
showing the better representation ability of vision transform-
ers. Although the ViTAEv2-S obtains the highest OA on the
Potsdam dataset, its mF1 is not as well as LANet [92]. From
Table VII, we can find that the scores of the “Car” category of
the selected models are worse than other methods. We suspect
that it may be because of the encoder-decoder structure and
the rough feature fusion strategy in the UperNet, where the
WANG et al. : EMPIRICAL STUDY OF REMOTE SENSING PRETRAINING 11
TABLE VI
RESULTS OF THE UPERNET SEGMENTATION MODEL WITH DIFFERENT BACKBONES AND SOTA METHODS ON THE TESTING SET OF THE POTSDAM
DATASET .
Method Backbone OA mF1F1 score per category
Imper. surf. Building Low veg. Tree Car
FCN [83] IMP-VGG-16 85.59 87.61 88.61 93.29 83.29 79.83 93.02
S-RA-FCN [84] IMP-VGG-16 88.59 90.17 91.33 94.70 86.81 83.47 94.52
Multi-filter CNN [85] IMP-VGG-16 90.65 85.23 90.94 96.98 76.32 73.37 88.55
FCN [83] IMP-ResNet-50 89.42 88.66 91.46 96.63 85.99 86.94 82.28
DANet [86] IMP-ResNet-50 89.72 89.14 91.61 96.44 86.11 88.04 83.54
PSPNet [87] IMP-ResNet-50 89.45 90.51 91.61 96.30 86.41 86.84 91.38
DeeplabV3+ [88] IMP-ResNet-50 89.74 90.94 92.35 96.77 85.22 86.79 93.58
ResT [89] IMP-ResNet-50 89.13 90.89 91.14 95.11 86.30 87.27 94.63
EaNet [90] IMP-ResNet-50 90.15 91.73 92.87 96.30 86.16 87.99 95.30 *
BotNet [91] IMP-ResNet-50 90.42 91.77 92.34 96.30 87.32 * 88.74 * 94.17
LANet [92] IMP-ResNet-50 90.84 91.95 * 93.05 97.19 * 87.30 88.04 94.19
UperNet IMP-ResNet-50 90.64 89.96 92.30 96.14 85.93 85.66 89.76
UperNet SeCo-ResNet-50 89.64 89.03 91.21 94.92 85.12 84.89 89.02
UperNet RSP-ResNet-50 90.61 89.94 92.42 96.15 85.75 85.49 89.87
UperNet IMP-Swin-T 91.17 90.60 92.94 96.66 86.54 85.87 90.98
UperNet RSP-Swin-T 90.78 90.03 92.65 96.35 86.02 85.39 89.75
UperNet IMP-ViTAEv2-S 91.60* 91.00 93.34* 96.84 87.28 86.38 91.18
UperNet RSP-ViTAEv2-S 91.21 90.64 93.05 96.62 86.62 85.89 91.01
TABLE VII
RESULTS OF THE UPERNET SEGMENTATION MODEL WITH DIFFERENT BACKBONES AND SOTA METHODS ON THE VALIDATION SET OF THE I SAID
DATASET .
Method Backbone mIOUIOU per category1
Ship ST BD TC BC GTF Bridge LV SV HC SP RA SBF Plane Harbor
FCN [83] IMP-VGG-16 41.7 51.7 22.9 26.4 74.8 30.2 27.9 8.2 49.3 37.0 0 30.7 51.9 52.1 62.9 42.0
UNet [93] - 39.2 49.0 0 36.5 78.6 22.9 5.5 7.5 49.9 35.6 0 38.0 46.5 9.7 74.7 45.6
DenseASPP [94] IMP-DenseNet-121 56.8 61.1 50.0 67.5 86.1 56.6 52.3 29.6 57.1 38.4 0 43.3 64.8 74.1 78.1 51.1
DenseUNet [95] IMP-DenseNet-121 58.7 66.1 50.4 76.1 86.2 57.7 49.5 33.9 54.7 46.2 0 45.1 65.9 71.9 82.2 54.6
Semantic FPN [96] IMP-ResNet-50 59.3 63.7 59.5 71.8 86.6 57.8 51.6 34.0 59.2 45.1 0 46.4 68.7 73.6 80.8 51.3
RefineNet [97] IMP-ResNet-50 60.2 63.8 58.6 72.3 85.3 61.1 52.8 32.6 58.2 42.4 23.0 43.4 65.6 74.4 79.9 51.1
PSPNet [87] IMP-ResNet-50 60.3 65.2 52.1 75.7 85.6 61.1 60.2 32.5 58.0 43.0 10.9 46.8 68.6 71.9 79.5 54.3
DeeplabV3 [98] IMP-ResNet-50 59.0 59.7 50.5 77.0 84.2 57.9 59.6 32.9 54.8 33.7 31.3 44.7 66.0 72.1 75.8 45.7
DeeplabV3+ [88] IMP-ResNet-50 60.8 63.9 52.5 72.8 84.9 56.5 58.9 32.2 59.1 42.9 31.4 46.1 67.7 72.9 79.8 52.6
EMANet [99] IMP-ResNet-50 55.4 63.1 68.4 66.2 82.7 56.0 18.8 42.1 58.2 41.0 33.4 38.9 46.9 46.4 78.5 47.5
ASP-OCNet [100] IMP-ResNet-50 40.2 47.3 40.2 44.4 65.0 24.1 29.9 27.1 46.3 13.6 10.3 34.6 37.9 41.4 68.1 38.0
DANet [86] IMP-ResNet-50 57.5 60.2 63.0 71.4 84.7 50.9 52.5 28.6 57.5 42.1 30.4 46.1 40.6 63.3 80.9 48.8
CCNet [101] IMP-ResNet-50 58.3 61.4 65.7 68.9 82.9 57.1 56.8 34.0 57.6 38.3 31.6 36.5 57.2 75.0 75.8 45.9
EncNet [102] IMP-ResNet-50 58.9 59.7 64.9 70.0 84.2 55.2 46.3 36.8 57.2 38.7 34.8 42.4 59.8 69.8 76.1 48.0
HRNet [103] IMP-HRNetW-18 61.5 65.9 68.9 74.0 86.9 59.4 61.5 33.8 62.1 46.9 14.9 44.2 52.9 75.6 81.7 52.2
RANet [104] IMP-ResNet-50 62.1 67.1 61.3 72.5 85.1 53.2 47.1 45.3* 60.1 49.3 38.1 41.8 70.5 58.8 83.1 55.6
AlignSeg [105] IMP-ResNet-50 62.1 67.4 68.9 76.2 86.2 62.1 52.0 28.7 60.7 50.3 31.2 45.7 56.2 71.2 82.9 54.8
OCR [106] IMP-HRNet-W48 62.6 67.8 70.7 73.6 87.9 63.4 47.7 33.1 61.4 49.6 30.4 48.4 59.5 72.8 83.3 53.3
HMANet [107] IMP-ResNet-50 62.6 65.4 70.9 74.7 88.7 60.5 54.6 29.0 59.7 50.3 32.6 51.4 62.9 70.2 83.8 51.9
FarSeg [108] IMP-ResNet-50 63.7 65.4 61.8 77.7 86.4 62.1 56.7 36.7 60.6 46.3 35.8 51.2 71.4* 72.5 82.0 53.9
FactSeg [109] IMP-ResNet-50 64.8 68.3 56.8 78.4* 88.9* 64.9* 54.6 36.3 62.7 49.5 42.7 51.5* 69.4 73.6 84.1 55.7
UperNet IMP-ResNet-50 61.9 65.9 73.9 68.1 70.7 57.3 52.5 39.2 61.2 48.8 34.3 44.5 62.1 76.8 83.8 52.2