High-Performance Siamese Network for Real-Time Tracking

Abstract

Target tracking algorithms based on deep learning have achieved good results in public datasets. Among them, the network tracking algorithm based on Siamese tracking has a high accuracy and fast speed, which has attracted significant attention. However, the Siamese tracker uses the AlexNet network as its backbone and the network layers are relatively shallow, so it does not make full use of the ability of the deep neural network. If only the backbones of target tracking are replaced, there will be no obvious improvement, such as in the cases of ResNet and Inception. Therefore, this paper designs a wider and deeper network structure. At a wider level, a mechanism that can adaptively adjust the receptive field (RF) size is designed. Firstly, multiple branches are divided by the split operator, and each branch has a different size of kernel corresponding to a different size of RF; then, the fuse operator is used to fuse the information of each branch to obtain the selection weights; and finally, according to the selection, the aggregation feature map is weighted. At a deeper level, a new kind of residual models is designed. The channel is simplified by pruning in order to improve the tracking speed. According to the above, a wider and deeper Siamese network was proposed in this paper. The experimental results show that the structure proposed in this paper achieves a good tracking effect and real-time performance on six kinds of datasets. The proposed tracker achieves an SUC and Prec of LaSOT of 0.569 and 0.571, respectively.

An Adaptive Kernels Layer for Deep Neural Networks Based on Spectral Analysis for Image Applications

“…Other works, such as [ 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 ], implemented what is known as Inception-like blocks, which use deep conventional networks, and the residual connections introduced in [ 26 ] to extract the outputs from each layer and concatenated them for the output, as shown in the example in Figure 4 , mimicking the operation using the Inception layer depth-wise by allowing the extracted feature at multiple receptive fields to be processed at the output layer. However, while this approach can be suitable for certain applications such as classification, a drop in the spatial features accumulates as we move deeper, diminishing the spatial accuracy of the larger LRF values, as illustrated in Figure 5 , where it can be seen that a bias towards features at the centre starts to increase, impairing the capability of the layer to accurately position where the feature is located and decreasing its efficiency in applications such as object detection.…”

Improved prostate cancer diagnosis using a modified ResNet50-based deep learning architecture

“…In this paper, we present a Prostate Cancer Detection Model (PCDM) depends on a modified ReseNet, a faster R-CNN mask, and dual optimizers (Adam and SGD) for detecting prostate cancer that applied on Prostate Cancer dataset [ 11 – 14 ]. PCDM model combines the power of DL with the accuracy of traditional methods to provide an effective method for detecting prostate cancer [ 15 , 16 ].…”