Antidecay LSTM for Siamese Tracking With Adversarial Learning.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Visual tracking is crucial in computer vision but faces challenges due to appearance changes.
• Existing Convolutional Neural Network (CNN)-based trackers often fail to capture long-term appearance variations.
• Recurrent Neural Network (RNN) approaches suffer from feature decay, degrading performance.

Purpose of the Study:

• To propose an anti-decay long short-term memory (AD-LSTM) for robust Siamese visual tracking.
• To enhance the standard LSTM architecture for improved temporal feature representation in tracking.
• To develop an adversarial learning algorithm for optimizing the AD-LSTM and Siamese network.

Main Methods:

• Implemented an AD-LSTM by replacing fully connected layers with convolutional layers for spatial feature extraction.
• Improved the LSTM cell architecture to minimize information decay over time.
• Utilized adversarial learning to train the AD-LSTM and generate robust target feature maps.

Main Results:

• The proposed AD-LSTM effectively preserves target appearance information in the temporal domain.
• Adversarial learning improved the accuracy of response map generation and robustness of feature maps.
• The tracker demonstrated superior performance compared to state-of-the-art methods on six benchmark datasets.

Conclusions:

• The AD-LSTM offers a significant advancement in visual tracking by addressing feature decay.
• Adversarial learning enhances the stability and accuracy of Siamese trackers.
• The proposed method achieves state-of-the-art results on challenging visual tracking benchmarks.