Feature Distilled Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Feature extraction is crucial for visual tracking accuracy and speed.
• Very deep Convolutional Neural Networks (CNNs) offer strong feature extraction but are computationally expensive for real-time applications.
• Model compression is explored to create efficient shallow networks for visual tracking.

Purpose of the Study:

• To develop a small, fast, and accurate model for visual tracking using model compression techniques.
• To address the computational complexity of deep CNNs in real-time visual tracking.
• To improve the efficiency of feature extraction for robust object tracking.

Main Methods:

• A small Feature Distilled Network (FDN) was proposed, mimicking intermediate representations of deeper networks.
• A shift-and-stitch method was introduced to reduce arithmetic operations while preserving feature map resolution.
• A scale adaptive discriminative correlation filter was learned on distilled features for scale variation handling.

Main Results:

• The proposed FDN achieved a 5x speed-up compared to deeper networks.
• The approach demonstrated competitive performance against state-of-the-art deep trackers on benchmark datasets.
• Rich visual features were extracted efficiently by the shallow FDN model.

Conclusions:

• The Feature Distilled Network (FDN) offers an effective solution for real-time visual tracking by balancing speed and accuracy.
• Model compression techniques, including feature distillation and shift-and-stitch, are viable for creating efficient deep learning models.
• The proposed method successfully handles scale variations in object tracking with improved computational efficiency.