High-Performance Siamese Network for Real-Time Tracking

Area of Science:

• Computer Vision
• Deep Learning
• Object Tracking

Background:

• Deep learning target tracking algorithms show promise on public datasets.
• Siamese network trackers offer high accuracy and speed but have limitations due to shallow backbones like AlexNet.
• Replacing backbones with ResNet or Inception does not significantly improve performance.

Purpose of the Study:

• To design a wider and deeper Siamese network for improved target tracking.
• To enhance the feature extraction capabilities of Siamese trackers.
• To achieve real-time performance with high accuracy.

Main Methods:

• A wider network level incorporates an adaptive receptive field (RF) mechanism using split and fuse operators.
• A deeper network level utilizes a novel residual model with channel pruning for speed optimization.
• The proposed architecture is a wider and deeper Siamese network.

Main Results:

• The proposed Siamese network demonstrates a good tracking effect and real-time performance across six datasets.
• Achieved a Success Rate (SUC) of 0.569 and Precision (Prec) of 0.571 on the LaSOT dataset.
• Outperforms existing Siamese trackers by leveraging a more advanced network architecture.

Conclusions:

• The developed wider and deeper Siamese network effectively enhances target tracking.
• The adaptive RF and pruned residual models contribute to improved accuracy and speed.
• The proposed tracker achieves state-of-the-art results, balancing performance and efficiency.