High-Performance Long-Term Tracking with Meta-Updater

Abstract: Long-term visual tracking has drawn increasing attention because it is much closer to practical applications than short-term tracking. Most top-ranked long-term trackers adopt the offline-trained Siamese architectures, thus, they cannot benefit from great progress of short-term trackers with online update. However, it is quite risky to straightforwardly introduce online-update-based trackers to solve the long-term problem, due to long-term uncertain and noisy observations. In this work, we propose a novel offline-trained Meta-Updater to address an important but unsolved problem: Is the tracker ready for updating in the current frame? The proposed meta-updater can effectively integrate geometric, discriminative, and appearance cues in a sequential manner, and then mine the sequential information with a designed cascaded LSTM module. Our meta-updater learns a binary output to guide the tracker's update and can be easily embedded into different trackers. This work also introduces a long-term tracking framework consisting of an online local tracker, an online verifier, a SiamRPN-based re-detector, and our meta-updater. Numerous experimental results on the VOT2018LT, VOT2019LT, OxUvALT, TLP, and LaSOT benchmarks show that our tracker performs remarkably better than other competing algorithms. Our project is available on the website: https://github.com/Daikenan/LTMU.

• The paper introduces a Meta-Updater that optimally decides when to update tracking models, addressing long-term tracking challenges.
• It integrates online tracker strengths using a cascaded LSTM model to sequentially combine geometric, discriminative, and appearance cues.
• Experimental results on benchmarks like VOT2018LT and OxUvALT demonstrate significantly improved F-scores and robust tracking performance.

High-Performance Long-Term Tracking with Meta-Updater

In the field of visual tracking, the transition from short-term to long-term tracking is of significant interest due to its closer alignment with practical applications. Traditional short-term tracking approaches are insufficient for handling the complex scenarios commonly encountered in long-term tracking, which involve frequent target disappearances and reappearances. This has led to a growing focus on developing robust long-term tracking algorithms. The paper "High-Performance Long-Term Tracking with Meta-Updater" offers a novel approach that addresses key challenges in this domain through the introduction of an offline-trained Meta-Updater.

Overview

Most existing long-term trackers rely on offline-trained Siamese architectures, which limits their ability to leverage advancements in short-term tracking methodologies that employ online updates. However, integrating online-update-based trackers into long-term scenarios is risky due to the potential contamination of the model with noisy samples arising from uncertain observations over extended periods. This paper proposes a Meta-Updater that determines when the tracker is ready to update in the current frame, effectively addressing this issue.

Meta-Updater Design

The Meta-Updater is an offline-trained module that integrates geometric, discriminative, and appearance cues sequentially. It employs a cascaded LSTM model to mine sequential information and provides a binary output to guide the tracker's update process. The Meta-Updater's design allows it to be easily embedded into various tracking frameworks, enhancing performance by minimizing the risk of online update contamination.

Long-Term Tracking Framework

The paper introduces a comprehensive long-term tracking framework comprising an online local tracker, an online verifier, a SiamRPN-based re-detector, and the Meta-Updater. This architecture benefits from the strengths of online-updated short-term trackers while mitigating associated risks. The experimental results demonstrate superior performance, with the proposed tracker achieving higher F-scores on long-term benchmarks such as VOT2018LT, VOT2019LT, OxUvALT, TLP, and LaSOT.

Experimental Findings

The experimental evaluation showcases the efficacy of the Meta-Updater and the long-term tracking framework. On the VOT2018LT dataset, the tracker achieved a remarkable F-score of 0.690, outperforming competing approaches like SiamRPN++ and SPLT. Similarly, in the VOT2019LT evaluation, it maintained the leading position with an F-score of 0.697. The OxUvALT dataset further validated the performance with a MaxGM of 0.751, demonstrating both high true positive rate and true negative rate.

Implications and Future Work

The incorporation of the Meta-Updater addresses critical challenges in long-term tracking by providing a robust mechanism for update decision-making, thus enhancing tracking reliability and accuracy. The demonstrated generalization ability of the Meta-Updater across various tracking algorithms suggests its potential for broader applications in the field. Future research could explore the integration of more sophisticated models and the expansion of the Meta-Updater's applicability to other domains requiring long-term observation and analysis.

By offering an innovative solution to the inherent difficulties of long-term tracking, this paper contributes meaningfully to advancing visual tracking technology, enabling trackers to operate effectively in complex, real-world scenarios.