SiamMixer: A Lightweight and Hardware-Friendly Visual Object-Tracking Network.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Siamese networks are widely used for visual object tracking.
• Existing research prioritizes accuracy over efficiency, limiting deployment on resource-constrained devices.
• Runtime memory cost is a critical but often overlooked factor in Siamese network design.

Purpose of the Study:

• To introduce SiamMixer, a lightweight and hardware-friendly visual object tracking network.
• To address parameter redundancy and reduce computational and memory overhead in Siamese networks.
• To enable efficient deployment of Siamese-network-based trackers on edge devices.

Main Methods:

• Patch-by-patch inference in shallow layers to minimize memory usage.
• Merging and global encoding of feature maps in deep layers for enhanced accuracy.
• Replacing all activation functions with ReLU for reduced computational cost.

Main Results:

• SiamMixer achieves comparable accuracy to larger trackers with significantly fewer parameters (286 kB).
• Demonstrates minimal extra memory usage for feature maps (196 kB).
• ReLU activation function contributes to reduced deployment costs on mobile devices.

Conclusions:

• SiamMixer offers a novel approach to efficient visual object tracking.
• The proposed methods effectively reduce memory and computational requirements.
• SiamMixer is suitable for deployment on edge and mobile devices due to its lightweight design.