Dynamical Hyperparameter Optimization via Deep Reinforcement Learning in Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Hyperparameter selection is crucial for deep learning-based visual object tracking performance.
• Current hyperparameter optimization methods often use generic ranges, failing to adapt to video-specific needs.
• Optimizing hyperparameters for individual video sequences presents significant challenges.

Purpose of the Study:

• To propose a novel dynamic hyperparameter optimization method for visual object tracking.
• To address the limitations of existing methods in adapting hyperparameters to specific video inputs.
• To enhance the performance and generalizability of deep learning-based trackers.

Main Methods:

• A novel dynamic hyperparameter optimization method using an action-prediction network based on continuous deep Q-learning.
• An efficient heuristic strategy to manage high-dimensional state spaces in object tracking.
• Application of the proposed algorithm to Siamese-based and correlation-filter-based trackers.

Main Results:

• The proposed method adaptively optimizes hyperparameters for individual video sequences.
• The approach successfully handles complex observation spaces in object tracking.
• Improved performance of Siamese-based and correlation-filter-based trackers demonstrated on popular benchmarks.

Conclusions:

• The proposed dynamic hyperparameter optimization method offers significant improvements in visual object tracking.
• The method demonstrates generalizability across different tracking algorithms.
• This work provides a more effective approach to hyperparameter tuning in video analysis.