text stringlengths 0 820 |
|---|
2) Implementation Detail and Experimental Setting: The |
training settings are the same as previous experiments. The |
training epoch and batch size are set to 200 and 64, respec- |
tively. These experiments are conducted on a single V100 |
GPU. Following [34], five settings of these three datasets are |
adopted to comprehensively evaluate the RS pretrained modelsand make the experiments become convincible, including |
UCM (8:2), AID (2:8), AID (5:5), NWPU-RESISC (1:9), and |
NWPU-RESISC (2:8). Note the m:nmeans 10m%samples |
are used for training, while the others form the testing set. |
Similar to the previous section, the images in each category |
are proportionally divided into two groups that are separately |
used for training and evaluation, respectively. Besides the |
above three backbones we selected, the ImageNet pretrained |
ResNet-50 and the ResNet-50 pretrained by SeCo [58] – an RS |
self-supervised method considering seasonal variation, are also |
adopted for a fair comparison. When implementing finetuning |
on each scene recognition task, only the neuron number of |
the last linear layer is changed to match the categories of |
the target dataset. The overall accuracy (OA), which is the |
most commonly used criterion in the aerial scene recognition |
community by counting the proportion of the correct classified |
images relative to all images in the testing set, is utilized in the |
experiments. The models are repeatedly trained and evaluated |
five times at each setting, and the average value µand standard |
deviationσof the results in different trials are recorded as |
µ±σ. |
3) Experimental Results: Quantitative Results and Anal- |
yses: Table V presents the results of the above selected |
backbones pretrained using different methods and other SOTA |
methods. Since this research only focuses on the pretraining |
of deep networks, especially the vision transformers. We only |
lists the DL based aerial scene recognition methods. For |
convenience, the “IMP” and “RSP” are used to represent |
“ImageNet Pretraining” and “Remote Sensing Pretraining”, |
respectively. It can be seen that the methods are split into five |
groups. The first group is the methods that adopt ResNet-50 |
as the backbone network, where the ResNet-50 is initialized |
by the ImageNet pretrained weights. This group can be used |
to compare with the third group. The second group includes |
the recent existing advanced methods whose backbone is |
other popular networks except for ResNet-50, such as the |
ImageNet pretrained VGG-16, ResNet-101, DenseNet-121, |
and so on. Then, the ResNet-50, Swin-T, and ViTAEv2-S |
networks, whose pretrained weights are obtained by IMP, RSP, |
or SeCo, form the last three groups, respectively. In addition, |
it should be noted that besides the network types, the weights |
pretrained for different epochs are also considered. The bold |
fonts in the last three groups mean the best results in each |
group, while “ *” denotes the best among all models (same |
meanings in other tasks). |
On the foundation of ImageNet pretrained ResNet-50, many |
methods are developed, which have been shown in the first |
group. Among these methods, many flexible and advanced |
modules have been explored. For example, the attention mech- |
anisms (CBAM [35], EAM [71], MBLANet [34]), where |
specific channels or spatial positions of the features are high- |
lighted, and multiscale features (F2BRBM [72] and GRMANet |
[73]), where the intermediate features are also employed. |
In addition, the self-distillation technology combined with |
specially designed loss functions (ESD-MBENet [25]) and |
the multibranch siamese networks (IDCCP [74]) have also |
been applied. While in the second group, the more diverse |
frameworks with various backbones are presented. Besides |
8 JOURNAL OF L ATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 |
TABLE V |
RESULTS OF THE SELECTED MODELS AND SOTA METHODS ON THE THREE SCENE RECOGNITION DATASETS UNDER DIFFERENT SETTINGS . THE BOLD |
FONTS IN THE LAST THREE GROUPS MEAN THE BEST RESULTS ,WHILE “*”DENOTES THE BEST AMONG ALL MODELS . |
Model Publication UCM (8:2) AID (2:8) AID (5:5) NWPU-RESISC (1:9) NWPU-RESISC (2:8) |
CBAM [35] ECCV2018 99.04±0.23 94.66±0.39 96.90±0.04 92.10±0.04 94.26±0.12 |
EAM (IMP-ResNet-50) [71] GRSL2021 98.98±0.37 93.64±0.25 96.62±0.13 90.87±0.15 93.51±0.12 |
F2BRBM (IMP-ResNet-50) [72] JSTARS2021 99.58±0.23 96.05±0.31 96.97±0.22 92.74±0.23 94.87±0.15 |
MBLANet (IMP-ResNet-50) [34] TIP2021 99.64±0.12 95.60±0.17 97.14±0.13 92.32±0.15 94.66±0.11 |
GRMANet (IMP-ResNet-50) [73] TGRS2021 99.19±0.10 95.43±0.32 97.39±0.24 93.19±0.42 94.72±0.25 |
IDCCP (IMP-ResNet-50) [74] TGRS2021 99.05±0.20 94.80±0.18 96.95±0.13 91.55±0.16 93.76±0.12 |
ESD-MBENet-v1 (IMP-ResNet-50) [25] TGRS2021 99.81±0.10 96.00±0.15 98.54±0.17 92.50±0.22 95.58±0.08 |
ESD-MBENet-v2 (IMP-ResNet-50) [25] TGRS2021 99.86±0.12 95.81±0.24 98.66±0.20 93.03±0.11 95.24±0.23 |
ARCNet (IMP-VGG-16) [75] TGRS2019 99.12±0.40 88.75±0.40 93.10±0.55 — — |
SCCov (IMP-VGG-16) [76] TNNLS2019 99.05±0.25 93.12±0.25 96.10±0.16 89.30±0.35 92.10±0.25 |
KFBNet (IMP-DenseNet-121) [77] TGRS2020 99.88±0.12 95.50±0.27 97.40±0.10 93.08±0.14 95.11±0.10 |
GBNet (IMP-VGG-16) [24] TGRS2020 98.57±0.48 92.20±0.23 95.48±0.12 — — |
MG-CAP (IMP-VGG-16) [78] TIP2020 99.00±0.10 93.34±0.18 96.12±0.12 90.83±0.12 92.95±0.13 |
EAM (IMP-ResNet-101) [71] GRSL2021 99.21±0.26 94.26±0.11 97.06±0.19 91.91±0.22 94.29±0.09 |
IMP-ViT-B [44] ICLR2021 99.28±0.23 93.81±0.21 96.08±0.14 90.96±0.08 93.96±0.17 |
MSANet (IMP-ResNet-101) [79] JSTARS2021 98.96±0.21 93.53±0.21 96.01±0.43 90.38±0.17 93.52±0.21 |
CTNet (IMP-MobileNet-V2+IMP-ViT-B) [52] GRSL2021 — 96.25±0.10 97.70±0.11 93.90±0.14 95.40±0.15 |
LSENet (IMP-VGG-16) [32] TIP2021 99.78±0.18 94.41±0.16 96.36±0.19 92.23±0.14 93.34±0.15 |
DFAGCN (IMP-VGG-16) [23] TNNLS2021 98.48±0.42 — 94.88±0.22 — 89.29±0.28 |
MGML-FENet (IMP-DenseNet-121) [26] TNNLS2021 99.86±0.12 96.45±0.18 98.60±0.04 92.91±0.22 95.39±0.08 |
ESD-MBENet-v1 (IMP-DenseNet-121) [25] TGRS2021 99.86±0.12 96.20±0.15 98.85±0.13* 93.24±0.15 95.50±0.09 |
ESD-MBENet-v2 (IMP-DenseNet-121) [25] TGRS2021 99.81±0.10 96.39±0.21 98.40±0.23 93.05±0.18 95.36±0.14 |
IMP-ResNet-50 [12] CVPR2016 98.81±0.23 94.67±0.15 95.74±0.10 90.09±0.13 94.10±0.15 |
SeCo-ResNet-50 [58] ICCV2021 97.86±0.23 93.47±0.08 95.99±0.13 89.64±0.17 92.91±0.13 |
RSP-ResNet-50-E40 Ours 99.43±0.24 95.88±0.07 97.29±0.07 92.86±0.09 94.40±0.05 |
RSP-ResNet-50-E120 Ours 99.52±0.15 96.60±0.04 97.78±0.08 93.76±0.03 94.97±0.07 |
RSP-ResNet-50-E300 Ours 99.48±0.10 96.81±0.03 97.89±0.08 93.93±0.10 95.02±0.06 |
IMP-Swin-T [13] ICCV2021 99.62±0.19 96.55±0.03 98.10±0.06 92.73±0.09 94.70±0.10 |
RSP-Swin-T-E40 Ours 99.24±0.18 95.95±0.06 97.52±0.04 91.22±0.18 93.30±0.08 |
RSP-Swin-T-E120 Ours 99.52±0.00 96.73±0.07 98.20±0.02 92.02±0.14 93.84±0.07 |
RSP-Swin-T-E300 Ours 99.52±0.00 96.83±0.08 98.30±0.04 93.02±0.12 94.51±0.05 |
IMP-ViTAEv2-S [29] arXiv2022 99.71±0.10 96.61±0.07 98.08±0.03 93.90±0.07 95.29±0.12 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.