LasHeR: A Large-Scale High-Diversity Benchmark for RGBT Tracking

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• RGBT tracking is a rapidly growing area in computer vision.
• Existing datasets lack the scale and diversity needed for robust RGBT tracker development and evaluation.

Purpose of the Study:

• To introduce LasHeR, a novel large-scale, high-diversity benchmark dataset for short-term RGBT tracking.
• To provide a comprehensive resource for training and evaluating RGBT tracking algorithms.

Main Methods:

• The LasHeR dataset comprises 1224 visible and thermal infrared video pairs, totaling over 730,000 frame pairs.
• Each frame pair is spatially aligned and meticulously annotated with bounding boxes.
• The dataset exhibits high diversity in object categories, camera viewpoints, scene complexity, and environmental factors.

Main Results:

• A comprehensive performance evaluation of 12 RGBT tracking algorithms was conducted on the LasHeR dataset.
• Detailed analysis of algorithm performance across diverse scenarios is presented.
• An unaligned version of LasHeR is released to promote research in alignment-free RGBT tracking.

Conclusions:

• LasHeR addresses the critical need for a large-scale, diverse benchmark in RGBT tracking.
• The dataset facilitates advancements in training deep RGBT trackers and evaluating their real-world applicability.
• The release of LasHeR and its unaligned variant is expected to spur further innovation in RGBT tracking research.