Convolutional Regression for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Discriminatively learned correlation filters (DCF) are popular in visual object tracking.
• DCF trackers utilize many samples for ridge regression, but synthetic samples can reduce robustness.

Purpose of the Study:

• To propose a new approach for visual tracking using a single convolutional layer and gradient descent (GD).
• To address the limitations of synthetic samples in DCF by incorporating real image samples.

Main Methods:

• Optimizing a one-channel output convolution layer with gradient descent (GD) instead of closed-form linear regression.
• Setting the convolution layer kernel size to the object size for incorporating real samples.
• Developing a novel objective function to handle negative and positive samples effectively.

Main Results:

• The proposed algorithm achieved outstanding performance on OTB-100, OTB-50, TempleColor, and VOT-2016 datasets.
• The method outperformed most existing DCF-based algorithms in extensive experiments.

Conclusions:

• The proposed GD approach with a single convolutional layer offers a robust and effective alternative for visual object tracking.
• The novel objective function successfully mitigates issues with negative samples, enhancing tracking accuracy.