SMTrack: State-Aware Mamba for Efficient Temporal Modeling in Visual Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual tracking aims to estimate object states in videos, facing challenges in dynamic environments.
• Conventional methods struggle with long-range temporal dependencies, requiring complex modules or high computational costs.
• State space models offer a promising alternative for temporal modeling.

Purpose of the Study:

• To introduce a novel temporal modeling paradigm for visual tracking.
• To develop a robust and computationally efficient tracking method.
• To overcome limitations of existing CNN and Transformer architectures in handling temporal information.

Main Methods:

• Proposing the State-aware Mamba Tracker (SMTrack), a new paradigm inspired by state space models.
• Implementing a selective state-aware space model with state-wise parameters for diverse temporal cue capture.
• Utilizing hidden state propagation for efficient frame-to-frame temporal interaction during tracking.

Main Results:

• SMTrack demonstrates effective modeling of long-range temporal dependencies.
• The method achieves robust tracking performance in dynamic scenarios.
• SMTrack offers linear computational complexity during training and reduced costs during tracking.

Conclusions:

• SMTrack provides a neat pipeline for visual tracking without complex customized modules.
• The proposed approach achieves promising performance with low computational costs.
• SMTrack advances the state-of-the-art in robust and efficient visual tracking.