Visual Tracking With Convolutional Random Vector Functional Link Network.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep neural networks excel in visual tracking but require large datasets like ImageNet for pretraining.
• Domain discrepancy between auxiliary data and target objects necessitates fine-tuning, increasing sensitivity to hyperparameters.
• Conventional pretraining and fine-tuning methods for visual tracking are complex and parameter-sensitive.

Purpose of the Study:

• To investigate the necessity of pretraining and fine-tuning in visual tracking.
• To propose a simplified visual tracking system that reduces reliance on auxiliary data.
• To introduce the convolutional random vector functional link (CRVFL) neural network for visual tracking.

Main Methods:

• Proposed a convolutional random vector functional link (CRVFL) neural network, integrating convolutional neural networks and random vector functional link networks.
• Kept convolutional layer parameters randomly initialized and fixed, learning only fully connected layer parameters.
• Developed an elegant update approach for the tracker.

Main Results:

• A single CRVFL model achieved 79.0% precision (20-pixel threshold) on a standard visual tracking benchmark without auxiliary data.
• An ensemble of CRVFL models achieved a leading result of 86.3% precision.
• Demonstrated effective visual tracking performance without extensive pretraining or fine-tuning.

Conclusions:

• The proposed CRVFL network simplifies visual tracking systems.
• Pretraining and fine-tuning via conventional backpropagation may not be essential for effective visual tracking.
• CRVFL offers a promising alternative for visual tracking, reducing data requirements and hyperparameter sensitivity.