CTT: CNN Meets Transformer for Tracking.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Siamese networks are prevalent in deep learning-based visual object tracking.
• Current methods often rely on feature pyramid networks (FPN) and cross-correlation for feature matching.
• Existing approaches may not sufficiently address the need for global and contextual dependencies in object tracking.

Purpose of the Study:

• To enhance visual object tracking by incorporating a novel residual transformer structure.
• To improve the ability of trackers to capture global and contextual information.
• To develop a more accurate and robust object tracking model.

Main Methods:

• Introduced a residual transformer structure with an encoder-decoder mechanism within the tracker's neck.
• The encoder facilitates interaction between low-level features for global attention.
• The decoder integrates global attention information into the head module, replacing cross-correlation.
• Incorporated spatial and channel attention in the target branch to boost performance.

Main Results:

• The proposed tracker, CTT, was evaluated on diverse benchmarks including GOT-10k, VOT2019, OTB-100, LaSOT, NfS, UAV123, and TrackingNet.
• CTT demonstrated competitive performance against state-of-the-art object tracking algorithms.
• The addition of spatial and channel attention improved accuracy and robustness with minimal computational cost.

Conclusions:

• The proposed encoder-decoder transformer structure effectively captures global dependencies in visual object tracking.
• CTT offers a promising advancement in deep learning-based object tracking, achieving state-of-the-art results.
• The model's enhanced accuracy and robustness make it suitable for various tracking applications.