Siamese network with a depthwise over-parameterized convolutional layer for visual tracking

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Visual tracking is crucial for applications like defense and security.
• Challenges include occlusion, fast motion, and background clutter.
• Siamese trackers with Convolutional Neural Networks (CNNs) offer good performance but underutilize spatial and semantic information, leading to drift.

Purpose of the Study:

• To address the limitations of existing visual trackers by improving target representation.
• To introduce a novel CNN feature extraction subnetwork utilizing Depthwise Over-parameterized Convolutional (DO-Conv) layers.
• To propose a new Siamese-based tracking algorithm, DOSiam, that effectively exploits spatial and semantic information.

Main Methods:

• Designed a CNN feature extraction subnetwork incorporating Depthwise Over-parameterized Convolutional (DO-Conv) layers.
• Introduced a joint convolution method combining conventional and depthwise convolutions.
• Developed the DOSiam algorithm within a Siamese framework, leveraging DO-Conv for enhanced feature extraction.

Main Results:

• The DO-Conv layer effectively extracts shallow spatial and deep semantic information while discarding background noise.
• DOSiam demonstrated superior tracking performance across five benchmark datasets (OTB2015, VOT2016, VOT2018, GOT-10k, VOT2019-RGBT(TIR)).
• Achieved real-time tracking speeds of 60 FPS, outperforming state-of-the-art trackers.

Conclusions:

• The proposed DO-Conv based Siamese tracker (DOSiam) significantly enhances target representation by integrating spatial and semantic information.
• DOSiam offers a robust solution for visual tracking challenges, achieving high accuracy and real-time performance.
• This work advances the field of visual tracking by improving feature extraction within Siamese frameworks.