text
stringlengths
0
820
dling of spatial interactions and leads to improved class
separation. Multiple clstokens have also been employed
in [11, 59]. However, while both studies use them as class
queries inputs to a decoder module, we introduce the clsto-
kens as an input to the encoder in order to collapse the time
dimension and obtain class-specific features.
3. Method
In this section we present the TSViT architecture in de-
tail. First, we give a brief overview of the ViT (section 3.1)
which provided inspiration for this work. In section 3.2 we
present our modified TSViT backbone, followed by our to-
kenization scheme (section 3.3), encoder (section 3.4) and
decoder (section 3.5) modules. Finally, in section 3.6, we
discuss several considerations behind the design of our po-
sition encoding scheme.
3.1. Primer on ViT
Inspired by the success of Transformers in natural lan-
guage processing tasks [64] the ViT [13] is an application
of the Transformer architecture to images with the fewest
Figure 2. Backbone architectures. (a) Transformer backbone,
(b)ViT architecture, (c)TSViT backbone employs additional cls
tokens (red), each responsible for predicting a single class.
possible modifications. Their framework involves the to-
kenization of a 2D image X∈RH×W×Cto a set of
patch tokens Z∈RN×dby splitting it into a sequence of
N=⌊H
h⌋⌊W
w⌋same-size and non-overlapping patches of
spatial extent (h×w)which are flattened into 1D tokens
xi∈RhwCand linearly projected into ddimensions. Over-
all, the process of token extraction is equivalent to the appli-
cation of 2D convolution with kernel size (h×w)at strides
(h, w)across respective dimensions. The extracted patches
are used to construct model inputs as follows:
Z0=concat (zcls,Z+P)∈RN+1×d(1)
A set of learned positional encoding vectors P∈RN×d,
added to Z, are employed to encode the absolute posi-
tion information of each token and break the permutation
invariance property of the subsequent Transformer layers.
A separate learned class ( cls) token zcls∈Rd[12] is
prepended to the linearly transformed and positionally aug-
mented patch tokens leading to a length N+ 1 sequence
of tokens Z0which are used as model inputs. The Trans-
former backbone consists of Lblocks of alternating lay-
ers of Multiheaded Self-Attention (MSA) [64] and residual
Multi-Layer Perceptron (MLP) (Fig.2(a)).
Yl=MSA (LN(Zl)) +Zl(2)
Zl+1=MLP (LN(Yl)) +Yl(3)
Prior to each layer, inputs are normalized following Layer-
norm (LN) [4]. MLP blocks consist of two layers of linear
projection followed by GELU non-linear activations [21].
In contrast to CNN architectures, in which spatial dimen-
sions are reduced while feature dimensions increase with
10420
Figure 3. SITS Tokenization . We embed each satellite image
independently following ViT [13]
layer depth, Transformers are isotropic in that all feature
maps Zl∈R1+N×dhave the same shape throughout the
network. After processing by the final layer L, all tokens
apart from the first one (the state of the clstoken) are dis-
carded and unormalized class probabilities are calculated by
processing this token via a MLP. A schematic representation
of the ViT architecture can be seen in Fig.2(b).
3.2. Backbone architecture
In the ViT architecture, the clstoken progressively re-
fines information gathered from all patch tokens to reach a
final global representation used to derive class probabilities.
Our TSViT backbone, shown in Fig.2(c), essentially fol-
lows from ViT, with modifications in the tokenization and
decoder layers. More specifically, we introduce K(equal to
the number of object classes) additional learnable clstokens
Zcls∈RK×d, compared to ViT which uses a single token.
Z0=concat (Zcls,Z+P)∈RN+K×d(4)
Without deviating from ViT, all clsand positionally aug-
mented patch tokens are concatenated and processed by the
Llayers of a Transformer encoder. After the final layer,
we discard all patch tokens and project each clstoken into
a scalar value. By concatenating these values we obtain a
length Kvector of unormalised class probabilities. This
design choice brings the following two benefits: 1) it in-
creases the capacity of the clstoken relative to the patch
tokens, allowing them to store more patterns to be used by
the MSA operation; introducing multiple clstokens can be
seen as equivalent to increasing the dimension of a single
clstoken to an integer multiple of the patch token dimen-
siondcls=kdpatch and split the clstoken into kseparate
subspaces prior to the MSA operation. In this way we can
increase the capacity of the clstokens while avoiding is-
sues such as the need for asymmetric MSA weight matri-ces for clsandpatch tokens, which would effectively more
than double our model’s parameter count. 2) it allows for
more controlled handling of the spatial interactions between
classes. By choosing k=Kand enforcing a bijective map-
ping from clstokens to class predictions, the state of each
clstoken becomes more focused to a specific class with net-
work depth. In TSViT we go a step further and explicitly
separate clstokens by class after processing with the tempo-
ral encoder to allow only same- cls-token interactions in the
spatial encoder. In section 3.4 we argue why this is a very
useful inductive bias for modelling spatial relationships in