LSOTB-TIR:A Large-Scale High-Diversity Thermal Infrared Object Tracking Benchmark (2008.00836v1)

Abstract: In this paper, we present a Large-Scale and high-diversity general Thermal InfraRed (TIR) Object Tracking Benchmark, called LSOTBTIR, which consists of an evaluation dataset and a training dataset with a total of 1,400 TIR sequences and more than 600K frames. We annotate the bounding box of objects in every frame of all sequences and generate over 730K bounding boxes in total. To the best of our knowledge, LSOTB-TIR is the largest and most diverse TIR object tracking benchmark to date. To evaluate a tracker on different attributes, we define 4 scenario attributes and 12 challenge attributes in the evaluation dataset. By releasing LSOTB-TIR, we encourage the community to develop deep learning based TIR trackers and evaluate them fairly and comprehensively. We evaluate and analyze more than 30 trackers on LSOTB-TIR to provide a series of baselines, and the results show that deep trackers achieve promising performance. Furthermore, we re-train several representative deep trackers on LSOTB-TIR, and their results demonstrate that the proposed training dataset significantly improves the performance of deep TIR trackers. Codes and dataset are available at https://github.com/QiaoLiuHit/LSOTB-TIR.

• The paper introduces LSOTB-TIR, a benchmark with 1,400 TIR sequences and over 600K frames annotated for robust evaluation of thermal tracking algorithms.
• It organizes data into 47 object classes with 4 scenario and 12 challenge attributes, enabling nuanced assessments of tracking performance.
• Evaluation of 30+ algorithms reveals that TIR-specific deep features significantly outperform traditional RGB-based methods in real-world applications.

Analyzing LSOTB-TIR: A Benchmark for Thermal Infrared Object Tracking

The paper "LSOTB-TIR: A Large-Scale High-Diversity Thermal Infrared Object Tracking Benchmark" introduces a comprehensive dataset designed specifically for evaluating thermal infrared object tracking algorithms. The LSOTB-TIR benchmark represents a significant effort in addressing limitations found in existing datasets, which include small scale, limited object categories, and lack of diversity in scenarios and challenges. By establishing a more representative and diverse dataset, LSOTB-TIR aims to facilitate the development and fair evaluation of deep learning-based thermal infrared (TIR) object tracking techniques.

Core Features of LSOTB-TIR

1. Scale and Diversity: LSOTB-TIR is distinguished by its large scale, comprising 1,400 TIR sequences amounting to over 600K frames with more than 730K annotated bounding boxes. The dataset encompasses 47 object classes across varied scenarios and challenges, making it both the largest and most diverse benchmark for TIR object tracking currently available.
2. Scenario and Challenge Attributes: The dataset is categorized into 4 scenario attributes and 12 challenge attributes, allowing for a nuanced evaluation of tracking algorithms. Challenges such as thermal crossover, distractors, and intensity variation are particularly noteworthy, given their prevalence and impact in real-world TIR tracking tasks.
3. Evaluation and Baselines: The paper evaluates over 30 tracking algorithms on LSOTB-TIR, providing baseline performances that highlight the effectiveness of deep learning-based trackers. The results indicate that deep trackers, especially those optimized specifically for TIR features, show promising performance improvements.

Implications and Analysis

The release of LSOTB-TIR is anticipated to propel advancements in TIR object tracking by providing a rigorous and varied benchmark for evaluation. The dataset's scale and diversity address the generalizability issues of previous benchmarks and encourage the development of more robust and accurate tracking models. The availability of scenario-specific subsets enables application-oriented performance evaluations, paving the way for tailored tracker improvement in different operational contexts.

Moreover, the paper demonstrates that TIR-specific deep learned features significantly outperform traditional RGB-based features in TIR tracking, underscoring the importance of developing models that cater specifically to the unique characteristics of infrared imagery, such as the lack of color information and subtle textural variances.

Future Directions

LSOTB-TIR catalyzes a pathway for future research to focus on:

• Developing More Adaptable Models: Given the dataset's scenario-specific data, research can delve into creating models capable of adapting dynamically to different environmental contexts.
• Leveraging TIR-Specific Features: Further exploration into TIR-specific features may continue to enhance the performance differential between RGB and TIR trackers.
• Real-World Applications: The dataset can drive progress in practical applications like surveillance, search and rescue, and autonomous vehicle navigation, where robust real-time TIR tracking is crucial.

In summary, the LSOTB-TIR benchmark serves as a cornerstone for forward-looking developments in thermal infrared object tracking by providing a detailed and diverse dataset that refines the evaluation process and fosters innovation in tracker algorithms.