Deformable Parts Correlation Filters for Robust Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Object Tracking

Background:

• Deformable parts models offer potential for tracking nonrigid objects but often underperform holistic methods due to complexity.
• Existing models simplify constellation topology for tractability, limiting their accuracy in object localization.

Purpose of the Study:

• To develop a novel formulation of deformable parts models using correlation filters for improved object tracking.
• To address limitations in current models by integrating geometric and visual constraints efficiently.

Main Methods:

• A new constellation model formulation employing correlation filters within a single convex cost function.
• Development of a two-level tracker: coarse localization via root filter and color model, refined by a deformable constellation of correlation filters.

Main Results:

• The proposed tracker achieves state-of-the-art performance on challenging benchmarks like OTB, VOT2014, and VOT2015.
• The method demonstrates real-time processing capabilities while maintaining high accuracy.

Conclusions:

• The novel correlation filter-based deformable constellation model significantly improves object tracking accuracy and efficiency.
• This approach offers a robust solution for real-time tracking of nonrigid objects, overcoming limitations of previous methods.