SiamMask: A Framework for Fast Online Object Tracking and Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual object tracking and video object segmentation are crucial tasks in computer vision.
• Existing methods often require separate models or struggle with real-time performance.
• Fully-convolutional Siamese approaches have shown promise but can be improved.

Purpose of the Study:

• To introduce SiamMask, a unified framework for real-time visual object tracking and video object segmentation.
• To enhance existing Siamese approaches by incorporating a binary segmentation task into the training process.
• To demonstrate the framework's efficiency and effectiveness on benchmark datasets.

Main Methods:

• SiamMask augments the loss function of fully-convolutional Siamese networks with a binary segmentation task during offline training.
• After training, SiamMask requires only a single bounding box for initialization.
• The framework simultaneously performs visual object tracking and segmentation at high frame-rates.
• Extension to multi-object tracking and segmentation is achieved via a cascaded multi-task model.

Main Results:

• SiamMask achieves high processing efficiency, operating at approximately 55 frames per second.
• The framework delivers real-time state-of-the-art performance on visual object tracking benchmarks.
• Competitive performance at high speeds is demonstrated on video object segmentation benchmarks.

Conclusions:

• SiamMask offers a simple yet effective unified approach for both visual object tracking and video object segmentation.
• The method achieves real-time performance with high accuracy and efficiency.
• The framework shows potential for extension to more complex multi-object scenarios.