LSOTB-TIR: A Large-Scale High-Diversity Thermal Infrared Single Object Tracking Benchmark

Area of Science:

• Computer Vision
• Machine Learning
• Sensor Fusion

Background:

• Thermal infrared (TIR) object tracking offers robust performance in low-visibility conditions, unlike visual tracking.
• The field of TIR object tracking lacks a standardized, large-scale benchmark for training and evaluation.
• This absence hinders the development and fair comparison of advanced tracking algorithms.

Purpose of the Study:

• To introduce LSOTB-TIR, a comprehensive benchmark dataset for thermal infrared single object tracking.
• To provide a unified platform for training and evaluating TIR object tracking algorithms.
• To encourage the development of deep learning-based TIR trackers and facilitate fair benchmarking.

Main Methods:

• Construction of LSOTB-TIR, comprising 1416 TIR sequences and over 643,000 frames.
• Annotation of bounding boxes for objects across all sequences, totaling over 770,000 annotations.
• Splitting the evaluation dataset into short-term and long-term subsets, with defined scenario and challenge attributes.

Main Results:

• LSOTB-TIR is the largest and most diverse TIR object tracking benchmark to date.
• Evaluation of 40 existing trackers on LSOTB-TIR provides baseline performance metrics.
• Retraining deep trackers on the LSOTB-TIR training set demonstrates significant performance improvements.

Conclusions:

• LSOTB-TIR serves as a crucial resource for advancing the field of thermal infrared object tracking.
• The benchmark facilitates comprehensive evaluation and promotes the development of more robust TIR tracking methods.
• The dataset and associated code are publicly available to foster community research and innovation.