IEEE Geoscience and Remote Sensing Magazine - December 2022 - 187

and inference speed. The following introduced anchor-free
trackers proved that a simpler framework contributes to an
easier training process and a faster inference speed.
SiamFC++
SiamFC++ [61] is based on SiamFC [52] with the crosscorrelation
operation between a template patch and search
patch. First, it divides the tracking task into the classification
and state estimation branches. Second, an anchorfree
offset regression strategy is designed to obtain the
position state of the target alone with the classification,
treated as a free prior knowledge method. Finally, a simple
yet effective quality assessment subbranch is added
to the classification branch to evaluate the prior spatial
score, which can refine the prediction result. In addition,
the tracker is provided with four kinds of famous backbones
(GoogLeNet, AlexNet, ShuffleNet, and TinyConv),
and these multiple choices are helpful for our comprehensive
analysis.
Ocean
Ocean [121] is a general anchor-free-based tracking framework.
First, an offline object-aware anchor-free network is
designed to predict the location and class directly. Second,
an online update network is used to learn the appearance
changes of the tracked object. Notice that it modifies the
ResNet-50 by decreasing the network parameters and increasing
receptive fields, improving the object localization
precision.
SiamCAR
SiamCAR [58] is based on an anchor-free network with
high precision. The framework consists of only two parts:
an original Siamese network and a classification-regression
network. The former makes the depth-wise crosscorrelation
for the multilayer convolution features from
search branch and template branch. The latter predicts
the foreground-background classification, score map
of center points (CPs), and regression values of the distances
from the CP to the object bounding box's border.
Thus, the framework is end to end, easily trained with
fewer hyperparameters.
SiamBAN
SiamBAN [59] is similar to SiamCAR; however, it uses a boxadaptive
head to predict the target position and the distance
to the four borders of the target bounding box. Two groups
of learnable weights are applied parallelly on the classification
and regression branches with the FC network (FCN).
ATTENTION-BASED TRACKERS
The attention mechanism has latent capacities to avoid
distractors and learns efficiently to distinguish the tracked
object from its background. Hence, it might have better performance
on tracking in densely similar objects, which is
commonly seen in SVs.
DECEMBER 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
SiamAttn
SiamAttn [60] is an attention-based tracking method that
realizes the spatial and the channel attention by self- and
cross-attention, respectively. The attentional backbone
features are generated by a Deformable Siamese Attention
(DSA) module to adapt the object shape efficiently. Then,
the depth-wise cross-correlation operation merges two attentional
features in the Siamese way. Finally, a series of
feature fusion and refinement blocks predicts the target's
bounding box with a mask.
CGACD
CGACD [63] is an attention-based tracking framework inspired
by the corner detection method, such as Gird R-CNN
[122]. Unlike the self- and cross-attention used in many attention-based
trackers, it realizes the spatial and the channel
attention by applying the pixel-wise and the channelwise
correlation between the target template and regions of
interest (ROIs), respectively. ROIs are obtained by SiamFC
[52]. Finally, the object's left-top and right-down corners
are obtained by the attentional feature map, and it has a
better performance in avoiding background distractors.
SiamGAT
SiamGAT [64] contains a specially designed attentionbased
graph module, which builds the feature relation between
the template and the search region in a part-to-part
way. It creates a target-aware area selection instead of a fullregion
template. Therefore, replacing the original depthwise
cross-correlation with this graph attention module
allows the Siamese tracking framework to make robust
tracking on different objects with multiple scales and aspect-ratio
variations. In addition, GoogLeNet is used as the
backbone for efficient feature extraction.
TRANSFORMER-BASED TRACKERS
The Transformer has recently shown incredible capabilities
in the CV field and is proficient in encoding consecutive sequences
and learning complex dependencies between input
sequences via self-attention. In SVs, the motion information
regarded as a crucial clue is a kind of temporal information,
which can be represented well by the Transformer.
TransX
TransX [65] is a transformer-assisted tracking framework
proposed to build temporal information relationships
among successive frames. Besides, it can be deployed on
existing Siamese or DCF-based trackers. Two typical transformer-assisted
trackers, named TrSiam and TrDiMP, are designed
based on SiamFC [52] and DiMP [110], respectively.
With the assistance of temporal information, the trackers
have made considerable progress in overall performance.
TransT
TransT [67] is another kind of transformer-based Siamese
tracker that replaces the cross-correlation operation with
187
IEEE Geoscience and Remote Sensing Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2022
