Tiny Object Tracking: A Large-Scale Dataset and a Baseline.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Tiny objects present significant challenges in computer vision tasks due to their weak appearance and features.
• Existing methods struggle with accurate tracking of small objects in complex scenarios.

Purpose of the Study:

• To advance research in tiny object tracking by introducing a comprehensive dataset and a robust baseline algorithm.
• To provide a benchmark for evaluating tiny object tracking performance under diverse conditions.

Main Methods:

• Creation of a large-scale video dataset with over 217K frames and high-quality bounding box annotations for 434 sequences.
• Inclusion of 12 challenge attributes to cover varied viewpoints and scene complexities for detailed analysis.
• Proposal of a novel multilevel knowledge distillation network (MKDNet) for enhanced feature representation, discrimination, and localization.

Main Results:

• MKDNet demonstrates superior performance compared to state-of-the-art methods on the proposed dataset.
• The multilevel knowledge distillation approach effectively enhances the tracking of tiny objects.
• The dataset facilitates attribute-based performance analysis, highlighting tracking challenges.

Conclusions:

• The developed dataset and MKDNet provide a strong foundation for tiny object tracking research.
• MKDNet offers an effective solution for improving the accuracy and robustness of small object tracking.
• The open-source availability of the dataset and code promotes further development in the field.