Video Text Tracking With a Spatio-Temporal Complementary Model.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Text tracking in videos is crucial for various applications.
• Existing tracking-by-detection methods struggle with complex scenarios like motion blur and similar text appearances, leading to incomplete trajectories and mis-associations.

Purpose of the Study:

• To propose a novel spatio-temporal complementary text tracking model.
• To enhance the accuracy and completeness of text trajectories in videos.

Main Methods:

• Utilizing a Siamese Complementary Module to exploit temporal continuity and reduce missed text detections.
• Integrating semantic and visual cues through a text similarity learning network for improved discriminative power.

Main Results:

• The proposed model effectively alleviates missed text detections, ensuring complete text trajectories.
• High discriminative power is achieved, preventing mis-associations between similar-looking text instances.
• State-of-the-art performance demonstrated on multiple public benchmarks.

Conclusions:

• The novel spatio-temporal complementary text tracking model significantly improves text tracking accuracy and robustness.
• The method offers a more reliable solution for complex video text tracking scenarios.