Appearance variation adaptation tracker using adversarial network.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep neural networks achieve high performance in visual object tracking.
• Overlapping initial target regions during training can lead to model overfitting.
• Improved model generalization is crucial for robust visual tracking.

Purpose of the Study:

• To propose the Appearance Variation Adaptation (AVA) tracker to enhance object tracking performance.
• To address model overfitting in deep learning-based trackers.
• To improve the generalization capability of visual trackers.

Main Methods:

• Developed an adversarial network with a generator and discriminator to align feature distributions.
• The generator learns an adaptation mask to counter the discriminator's classification.
• A gradient reverse layer enables end-to-end mini-max optimization.

Main Results:

• AVA achieved the highest Area Under Curve (AUC) and average precision on the OTB50 benchmark.
• Achieved top scores in precision on OTB100 and favorable results on VOT2016.
• Demonstrated superior performance compared to state-of-the-art trackers.

Conclusions:

• The proposed AVA tracker effectively improves object tracking by adapting appearance variations.
• Adversarial learning enhances model generalization, reducing overfitting issues.
• AVA demonstrates competitive and often superior performance on standard tracking benchmarks.