Learning Cross-Attention Discriminators via Alternating Time-Space Transformers for Visual Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) dominate visual tracking but struggle with long-range spatial dependencies.
• Transformer-assisted methods combine CNNs and Transformers, improving feature representation.
• Existing approaches often integrate Transformers with CNNs, limiting pure Transformer exploration.

Purpose of the Study:

• To propose a novel, pure Transformer-based visual tracking model.
• To overcome the limitations of convolutional operations in relating spatially distant information.
• To introduce a semi-Siamese architecture leveraging attention mechanisms exclusively.

Main Methods:

• Developed multistage alternating time-space Transformers (ATSTs) inspired by Vision Transformers (ViTs).
• Employed a time-space self-attention module for feature extraction backbone.
• Utilized a cross-attention discriminator for direct response map estimation without convolution or correlation filters.

Main Results:

• The ATST-based model demonstrated favorable results against state-of-the-art convolutional trackers.
• Achieved comparable performance to recent "CNN + Transformer" trackers on multiple benchmarks.
• Required significantly less training data compared to existing methods.

Conclusions:

• Pure Transformer architectures are viable and effective for visual tracking.
• The proposed ATST model offers a promising direction for efficient and robust visual tracking.
• Attention-based mechanisms can replace convolution for enhanced feature representation in tracking.