Target-aware transformer tracking with hard occlusion instance generation.

Area of Science:

• Computer Vision
• Deep Learning
• Object Tracking

Background:

• Transformer architectures are increasingly dominant in visual tracking, surpassing Siamese networks.
• Current transformer-based trackers struggle with occluded scenes due to difficulties recognizing incomplete target information.

Purpose of the Study:

• To propose a novel transformer tracking approach, TATT, designed to enhance performance in occluded visual tracking scenarios.
• To address the limitations of existing trackers in recognizing targets with incomplete appearance data.

Main Methods:

• Developed a target-aware transformer network with an encoder-decoder structure for feature interaction.
• Integrated a hard occlusion instance generation module using image similarity to create realistic occlusion scenarios.
• The target-aware transformer directly predicts target boundaries for tracking results.

Main Results:

• TATT demonstrates promising performance across five benchmarks (LaSOT, TrackingNet, Got10k, OTB100, UAV123).
• Achieved state-of-the-art AUC scores of 65.5% (partial occlusion) and 61.2% (full occlusion) on the LaSOT dataset.
• The tracker operates at approximately 41 frames per second on GPU.

Conclusions:

• TATT effectively improves visual tracking accuracy in challenging occluded environments.
• The proposed methods successfully handle incomplete target appearance information, outperforming existing approaches.
• TATT offers a robust and efficient solution for real-world visual tracking applications.