Geometric structure-constraint tracking with confident parts

• We construct an adaptive structure representation model of object.
• We infer state of object with confident parts of object.
• Confident parts are obtained by measuring appearance variance and deformation.
• Robust for tracking with multiple challenges especially deformation and occlusion.

Recent work has successfully tracked objects with single challenge by matching similar features or classifying positive samples. However, if appearance of object changes drastically when it simultaneously encounters multiple challenges especially occlusion, fast motion and deformation in real-world, it is hard to track object via similar appearance representation. Instead of finding suitable match, we propose an adaptable structure representation tracking (SRT) method to infer object state by exploiting confident parts of object based on mid-level appearance and structure constraints. The structure representation is graph-based and propagated by dynamic programming. The experimental results of our SRT demonstrate significant tracking performance especially in challenging environments.