text
stringlengths
0
820
“Type” is the specific attention types. The ViTAE-S adopts
the performer of T2T-ViT [63] in the first two reduction cells.
“L” means this reduction cell does not use PCM and attention.
“F” and “W” separately denote the MHSA and WMHSA in
ViT and Swin. At last, the “Depth” is the number of stacked
normal cells, which is also the Niin Figure 2.
C. Implementation
1) Determine the Pretraining Network: We first determine
the type of deep models to be used for RSP. To this end, from
the official training set, we construct a mini-training set and a
mini-evaluation set, which have 9,775 and 225 images, respec-
tively. Note the latter set is formed by randomly selecting 5
images from each category to balance the classes. For CNN,
the classic ResNet-50 [12] is employed. Since this research
mainly explores the performance of vision transformer models
with RSP, a series of typical vision transformer based networks
including DeiT-S [45], PVT-S [46], and Swin-T [13], are also
evaluated. The selection of a specific version is to guarantee
these models have a similar amount of parameters compared
with the ViTAE-S model. In addition, we also include ViT-B
[44] for reference since ViT is the most basic model of vision
transformers.
All models are trained with 300 epochs and batch size 16.
We adopt the AdamW optimizer where the momentum is set
to 0.9 and the weight decay is set to 0.05. The initial learning
rate is 1e-3, which is adjusted by the cosine scheduling
policy:currentlr=minlr+1
2(initiallr−minlr)(1 +
cos(iter
max iterπ)), whereminlris 5e-6. In addition, we set
the warming up epochs to 5, where the learning rate is set
to 5e-7. Following the typical IMP, the input image is resized
to 224 ×224 by randomly cropping during training, while
during testing, the image at the same size is obtained through
“center crop”. In addition, a series of data argumentations
including AutoAugment [65], Random Erasing [66], Mixup
[67], CutMix [68], and color jitter are applied to improve the
training performance. The top-1 accuracy and top-5 accuracy
are used as the evaluation metrics. In addition, all models are
implemented on a single NVIDIA Tesla V100 GPU, and the
results are shown in Table II.
6 JOURNAL OF L ATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
TABLE II
RESULTS OF DIFFERENT MODELS ON THE MINI -EVALUATION SET . THEY ARE TRAINED ON THE MINI -TRAINING SET FROM MILLION AID.
Model Acc@1 Acc@5 Param(M) Training Time (hh:mm:ss)
ResNet-50 [12] 80.8 96.1 23.6 16:42:58
DeiT-S [45] 72.9 92.9 21.7 16:57:39
ViT-B [44] 78.4 92.1 114.4 17:41:03
PVT-S [46] 80.0 94.1 23.9 18:06:42
Swin-T [13] 84.7 93.7 27.6 17:49:06
ViTAE-S [14] 85.9 96.9 22.8 24:00:35
ViTAEv2-S [29] 88.2 96.1 18.8 23:57:30
It can be seen that despite the ViT-B has the most param-
eters, it performs not better than the classic ResNet-50. The
DeiT-S performs the worst since we do not adopt the teacher
model. Because our task is pretraining using RS images,
obtaining the corresponding teacher model is our target instead
of the prerequisite. By introducing the design paradigm of the
feature pyramid, PVT-S improves the accuracy compared with
ViT-B. On this foundation, the original ViTAE-S further con-
siders the locality and scale-invariance modeling, which are the
inductive biases of conventional CNN models. However, it cost
much training time since the token number in the early RCs is
large, requiring more computations. The Swin-T addresses this
issue by restricting the MHSA in fixed windows and adopts
the image shifting to implicitly promote the communications
between windows. By taking the advantage of WMHSA, the
ViTAEv2-S achieves the best performance and it exceeds the
second place by 2.3% top-1 accuracy.
The procedure of model determination is shown as follows.
For the ViTAE models, we choose the strongest one to expect
good performance in downstream tasks such as the aerial
scene recognition when adopting RSP, i.e., the ViTAEv2-S.
For comparison, the ResNet-50 is selected as the representative
network in conventional CNN, and the RS pretrained ResNet-
50 can also provide a group of new CNN related baselines on a
series of aerial datasets. The DeiT-S and ViT-B are eliminated
because of the low accuracy and a large number of parameters,
and they are difficult to be transferred into downstream tasks
because of the design of stacking transformers. The Swin
can be regarded as building on the foundation of PVT by
substituting the global MHSA with the shiftable WMHSA.
Since the top-1 accuracy of Swin is larger than PVT, and
Swin-T requires less training time, we also choose Swin-T
in the subsequent experiments.
2) Obtain the Suitable Weights: After determining the
model candidates, we conduct the RSP to obtain the pretrained
weights. Concretely, maintaining the category balance, we
randomly choose 1,000 images in each category of the Million-
AID dataset to form the validation set that has 51,000 images,
achieving a similar volume compared with the ImageNet
validation set, which contains 50,000 images. The remaining
949,848 images are used for training. Although the number
of images and categories for RSP are less than those of the
ImageNet training set, it still can perform to be competitive
or even achieves SOTA results on aerial scene tasks, whose
details will be presented later.
To obtain suitable pretraining weights, we separately train
the ViTAEv2-S model under the configuration of different
epochs. The basic learning rate is 5e-4, and the batch sizeTABLE III