Adaptive Siamese Tracking With a Compact Latent Network.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Siamese-based trackers are popular for object tracking but suffer from failure cases due to insufficient training data.
• Existing methods struggle with adapting to new scenes and variations during tracking.

Purpose of the Study:

• To simplify Siamese-based trackers by viewing tracking as a classification problem.
• To address failure cases by incorporating sequence-specific information from initial frames.
• To improve adaptation speed and accuracy in dynamic tracking scenarios.

Main Methods:

• Re-framing the tracking task as a classification problem for intuitive analysis.
• Developing a compact latent network using statistics-based features from initial frames.
• Implementing a diverse sample mining strategy for enhanced network discrimination.
• Proposing a conditional updating strategy for efficient model adaptation during tracking.

Main Results:

• Successfully adapted three classical Siamese-based trackers (SiamRPN++, SiamFC, SiamBAN).
• Achieved superior performance in accuracy across six recent datasets.
• Maintained high running speeds, demonstrating efficiency.

Conclusions:

• The proposed method effectively enhances Siamese-based trackers by leveraging initial frame information.
• The approach improves both accuracy and speed, offering a robust solution for object tracking.
• This work provides a novel perspective and practical improvements for Siamese-based tracking algorithms.