Real-Time Correlation Tracking via Joint Model Compression and Transfer.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Correlation filters (CF) are efficient for visual tracking but struggle with computational demands when using deep features from Convolutional Neural Networks (CNNs).
• Existing deep CF trackers are too resource-intensive for deployment on mobile platforms with limited processing power, such as single-core CPUs.

Purpose of the Study:

• To develop a method for compressing and transferring knowledge from large, off-the-shelf CNN models to lightweight student networks for accelerated CF tracking.
• To enable real-time visual tracking on resource-constrained devices.

Main Methods:

• A knowledge distillation framework was employed, using a pre-trained CNN as a teacher network and distilling it into a lightweight student network.
• A fidelity loss was introduced to preserve the representation capability of the teacher network in the student network.
• A tracking loss was designed to adapt the student network for visual tracking tasks, and a background-aware online learning scheme was used for adaptive updates.

Main Results:

• The lightweight student network significantly speeds up state-of-the-art deep CF trackers to achieve real-time performance on single-core CPUs.
• Tracking accuracy was maintained at levels comparable to the original, more computationally expensive trackers.
• Experiments on six standard datasets validated the effectiveness of the proposed method.

Conclusions:

• Jointly compressing and transferring CNN models via knowledge distillation is an effective strategy for accelerating deep CF trackers.
• The proposed fidelity and tracking losses, combined with online adaptation, enable efficient and accurate visual tracking on mobile platforms.