Active Learning for Deep Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep Convolutional Neural Networks (CNNs) are effective for single target tracking.
• Training deep CNNs requires large, high-quality labeled datasets, which are costly and time-consuming to create.
• Manual data labeling presents a significant bottleneck in developing robust visual tracking systems.

Purpose of the Study:

• To propose an active learning method to reduce the labeling cost for training deep visual tracking models.
• To develop a tracker that achieves competitive performance while minimizing manual annotation efforts.
• To enhance the efficiency and practicality of deep learning-based visual tracking.

Main Methods:

• An active learning strategy is employed to select informative unlabeled samples for annotation.
• Multiframe collaboration ensures sample diversity, while nearest-neighbor discrimination screens for representativeness and quality.
• A Tversky loss function is utilized to improve bounding box estimation accuracy.

Main Results:

• The proposed active-learning-based tracker (ALT) demonstrates competitive tracking accuracy and speed.
• The method significantly reduces the required labeling cost compared to traditional approaches.
• ALT achieves state-of-the-art performance on seven challenging evaluation benchmarks.

Conclusions:

• Active learning offers an effective solution to the data labeling challenge in deep visual tracking.
• The proposed multiframe collaboration and nearest-neighbor discrimination methods enhance sample selection.
• The ALT tracker provides a practical and efficient approach for real-world visual tracking applications.