IoUformer: Pseudo-IoU prediction with transformer for visual tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Siamese tracking methods face inconsistencies where classification scores do not always yield optimal bounding box overlap with ground truth.
• This inconsistency negatively impacts overall tracking performance.

Purpose of the Study:

• To propose a novel Siamese tracking paradigm that directly predicts intersection over union (IoU) values to overcome bounding box estimation inconsistencies.
• To introduce IoUformer, a concise target state predictor that leverages transformer's long-range dependency modeling for improved tracking.

Main Methods:

• IoUformer extends transformer capabilities to model target-aware interactions between the target template and search region, as well as sub-region interactions.
• This approach unifies global semantic interaction and target state prediction by directly forecasting IoU values.
• The study offers network components and localization strategies for implementing the IoU-based Siamese tracking paradigm.

Main Results:

• IoUformer accurately predicts IoU values, showing a near-linear correlation with ground truth.
• The IoUformer-based tracker demonstrates promising results, even with reduced training data.
• The method serves as an effective refinement module for enhancing existing advanced trackers.

Conclusions:

• The proposed IoU-based Siamese tracking paradigm, powered by IoUformer, offers a robust solution to bounding box inconsistency issues.
• IoUformer's ability to predict reliable IoU values and its compatibility with existing trackers highlight its efficacy and portability.