EMAT: Efficient feature fusion network for visual tracking via optimized multi-head attention.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Transformer-based methods show promise in visual tracking but face computational challenges due to feature map partitioning.
• Traditional approaches often lead to high resource consumption and reduced efficiency in multi-head attention mechanisms.

Purpose of the Study:

• To design a novel feature fusion network with optimized multi-head attention for Transformer-based visual tracking.
• To enhance tracking accuracy and computational efficiency by reducing reliance on irrelevant background information.

Main Methods:

• Developed a new feature fusion network incorporating efficient multi-head self-attention and spatial reduction attention modules.
• Preprocesses input features to optimize multi-head attention calculations within an encoder-decoder Transformer architecture.

Main Results:

• The proposed EMAT tracker achieved superior performance on seven benchmark datasets (LaSOT, GOT-10k, TrackingNet, UAV123, VOT2018, NfS, VOT-RGBT2019).
• Key results include 89.0% precision on UAV123, 64.6% AUC on LaSOT, and 34.8% EAO on VOT-RGBT2019.
• The tracker operates at a real-time speed of approximately 35 FPS.

Conclusions:

• The novel feature fusion network and optimized attention modules significantly improve Transformer-based visual tracking.
• EMAT demonstrates state-of-the-art performance and real-time capabilities, outperforming existing advanced trackers.