Distilling Channels for Efficient Deep Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep trackers utilize pre-trained deep networks for object representation, excelling with generic objects.
• However, these networks are often too complex for specific moving object tracking, causing generalization issues and high computational demands.

Purpose of the Study:

• Introduce a novel framework, channel distillation, to improve deep tracker efficiency and performance.
• Adaptively select informative feature channels to overcome limitations of fixed-layer feature extraction.

Main Methods:

• Channel distillation unifies feature compression, response map generation, and model update into an energy minimization problem.
• It adaptively selects informative feature channels, reducing reliance on noisy ones and generalizing across networks.
• The framework was validated using discriminative correlation filter (DCF) and ECO trackers.

Main Results:

• Channel distillation effectively extracts relevant feature channels, enhancing tracking efficacy.
• The method significantly reduces the number of channels required, leading to lower computational and memory costs.
• The resulting deep tracker demonstrates improved accuracy, speed, and generalizability.

Conclusions:

• Channel distillation offers a general and effective approach to optimize deep trackers for moving object tracking.
• The framework successfully balances accuracy, speed, and resource efficiency.
• Experimental results confirm the broad applicability and effectiveness of channel distillation across benchmarks.