GOT-10k

GOT-10k provides a large-scale benchmark dataset for training and evaluating class-agnostic short-term generic object trackers across diverse real-world video scenarios.

Key Features:

• Extensive Dataset: GOT-10k comprises over 10,000 video segments with more than 1.5 million manually labeled bounding boxes for unified training and evaluation of deep tracking algorithms.
• Semantic Hierarchy (WordNet): The dataset leverages the WordNet structure to define classes, covering over 560 distinct moving object categories and 87 motion patterns.
• One-shot Evaluation Protocol: GOT-10k employs a one-shot protocol that enforces zero-overlap between training and test classes to measure tracker generalization.
• Additional Annotations: The dataset includes motion-class labels and object visible-ratio annotations to support motion-aware and occlusion-aware evaluation.
• Experimental Analysis: GOT-10k has been used to evaluate 39 typical tracking algorithms and their variants to analyze performance across diverse scenarios.

Scientific Applications:

• Generic Tracker Development: Training and evaluating class-agnostic short-term object trackers under realistic video conditions.
• Benchmarking Algorithms: Comparative analysis and benchmarking of tracking algorithms and variants across diverse object categories and motion patterns.
• Motion-aware Tracking: Research leveraging motion-class labels to develop and evaluate motion-sensitive tracking methods.
• Occlusion-aware Tracking: Research using visible-ratio annotations to study tracker robustness to occlusion and partial visibility.

Methodology:

Manual annotation of over 1.5 million bounding boxes; class selection via the WordNet semantic hierarchy; a one-shot evaluation protocol enforcing zero-overlap between training and test classes; inclusion of motion-class and visible-ratio annotations; empirical evaluation of trackers (39 algorithms and variants).