Robust Online Tracking With Meta-Updater

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object tracking models struggle with dramatic appearance changes in target and background.
• Online update schemes improve trackers but introduce noisy samples, risking accuracy.
• Existing methods lack a mechanism to assess the readiness for online updates.

Purpose of the Study:

• To propose a novel offline-trained Meta-Updater module for visual object tracking.
• To address the critical question of whether a tracker is ready for online updates in real-time.
• To enhance the robustness and accuracy of discriminative trackers facing appearance variations.

Main Methods:

• The Meta-Updater integrates geometric, discriminative, and appearance cues sequentially.
• A cascaded Long Short-Term Memory (LSTM) module mines sequential information.
• An additional local outlier factor strengthens appearance information within a new network design.

Main Results:

• The Meta-Updater was integrated into eight different online update trackers.
• Experiments were conducted on four long-term and two short-term tracking benchmarks.
• The proposed Meta-Updater demonstrated significant effectiveness and strong generalization capabilities.

Conclusions:

• The novel Meta-Updater effectively determines the readiness of trackers for online updates.
• This approach mitigates the risks associated with noisy observations in online learning.
• The Meta-Updater offers a generalizable solution to improve the performance of various visual trackers.