Generalized Pooling for Robust Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Sparse coding-based tracking often overlooks high-order statistics and correlations in feature representations.
• Existing methods rely on low-order statistics or extreme responses, limiting tracking performance.

Purpose of the Study:

• To develop a generalized feature pooling method for visual tracking.
• To enhance the discriminative power of feature representations by modeling sparse code distributions.
• To improve the accuracy and robustness of visual tracking algorithms.

Main Methods:

• Utilized probabilistic functions to model the statistical distribution of sparse codes.
• Introduced Fisher vectors for compact and discriminative sparse code representation.
• Employed local coordinate coding for target patch encoding.
• Used Gaussian mixture models to compute Fisher vectors.
• Trained semi-supervised linear kernel classifiers, updated online to handle tracking drift.

Main Results:

• The proposed Fisher vector-based feature pooling method significantly enhances visual tracking.
• The online updated classifiers effectively mitigate the drifting problem.
• Experimental results show superior performance compared to state-of-the-art tracking algorithms on challenging benchmarks.

Conclusions:

• The generalized feature pooling method offers a more robust and accurate approach to visual tracking.
• Fisher vectors provide a powerful tool for creating discriminative representations from sparse codes.
• The proposed method represents a significant advancement in sparse coding-based visual tracking.