Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
stages, it is unnecessary to partition the feature for WMHSA.
Thus, only the MHSAs in the first two stages are substituted by
WMHSA. It should be noticed that the adopted WMHSA does
not need to be shifted as the original implementation, since the
WANG et al. : EMPIRICAL STUDY OF REMOTE SENSING PRETRAINING 5
D-Conv D-Conv
ConcatGELU GELU
LN
MHSA
LN
FFNG-Conv
BN
SiLU
G-Conv
BN
SiLU Img2Seq
Img2Seq
Seq2Img…PRMPCMD-Conv D-Conv
ConcatGELU GELU
LN
WMHSA
LN
FFNImg2Seq
Img2Seq
Seq2Img…PRMPCM
G-Conv
SiLUG-Conv
BN
SiLU
G-Conv
BN
SiLU
G-Conv
LN
MHSA
LN
FFNG-Conv
BN
SiLUImg2Seq
Img2Seq
Seq2ImgPCM
G-Conv
SiLUG-Conv
BN
SiLULN
WMHSA
LN
FFNG-Conv
BN
SiLUImg2Seq
Img2Seq
Seq2ImgPCM
G-ConvG-Conv
BN
SiLU(a) (b)
(c) (d)
Fig. 3. The structures of different cells in ViTAE models. (a) and (c) are the
reduction cell and normal cell in the original ViTAE, while (b) and (d) are
the corresponded variants in the ViTAEv2.
WMHSA is conducted on the merged multiscale feature from
the pyramid reduction module (PRM), where different regions
have communicated with each other through the overlapped
receptive fields of the sliding dilated convolutions. Besides, it
is also not necessary to use relative positional encoding since
the convolutions already encode the positional information.
Additionally, the SiLU [64] in the last convolutional layer
of the parallel convolutional module (PCM) is also removed
to reduce nonlinearity. The structures and comparisons of
different cells in the original ViTAE and ViTAEv2 have been
shown in Figure 3. For reduction cell, normal cell, PRM and
PCM, readers can refer to [14] and [29] for more details.
In our implementation, we mainly evaluate the “small”
version of the original ViTAE, named ViTAE-S. In addition,
we also adopt the ViTAEv2-S model due to its excellent
representation ability and transferability to downstream tasks.
Table I lists the details of ViTAE-S and ViTAEv2-S. Here,
the length of the corresponded list equals the number of
stages. “Embedding Dim” means the encoding dimension in
PRM, while “Stage Dim” is the channel number of the feature
through the corresponding stage, which is useful for aligning
the related downstream task decoders. The “RC” and “NC”
separately represent the reduction cell and normal cell, where
“Head” is the head number in MHSA or WMHSA, “Group”
represents the number of group convolutions in PCM, andTABLE I
THE HYPERPARAMETER SETTINGS OF DIFFERENT “SMALL ”VERSION
VITAE MODELS . “P” DENOTES PERFORMER ATTENTION [63] WHILE “L”
MEANS THE REDUCTION CELL HAS NO PCM AND ATTENTION . “F”
DENOTES THE ORIGINAL MHSA WHILE “W” DENOTES THE WMHSA.
Network ViTAE-S [14] ViTAEv2-S [29]
Stage 3 4
Downsampling Ratio [4, 2, 2] [4, 2, 2, 2]
Embedding Dim [64, 64, 192] [64, 64, 128, 256]
Stage Dim [96, 192, 384] [64, 128, 256, 512]
RCHead [1, 1, 1] [1, 1, 2, 4]
Group [1, 1, 1] [1, 16, 32, 64]
Type [P, P, L] [W, W, F, F]
NCHead [1, 1, 6] [1, 2, 4, 8]
Group [1, 1, 96] [1, 32, 64, 128]
Type [F, F, F] [W, W, F, F]
Depth [0, 0, 14] [2, 2, 8, 2]