Discriminative Tracking Using Tensor Pooling.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Effective global representation from local descriptors is crucial for vision tasks.
• Local sparse representation offers discriminative and robust features for visual tracking.
• Existing methods often convert sparse codes to vectors, losing structural information.

Purpose of the Study:

• To propose a novel method for visual tracking using sparse coding tensors.
• To leverage the intrinsic structural information preserved in tensor representations.
• To improve robustness against noise and drifting compared to vector-based methods.

Main Methods:

• Representing target templates and candidates using sparse coding tensors.
• Building an appearance model through incremental learning on these tensors.
• Developing a discriminative framework to enhance robustness.

Main Results:

• The proposed tensor-based sparse coding method outperforms state-of-the-art trackers.
• Preserving tensor structure provides richer target appearance information.
• The discriminative framework effectively combats drifting and environmental noise.

Conclusions:

• Sparse coding tensors offer a superior representation for visual tracking compared to traditional vector pooling.
• The proposed method demonstrates enhanced performance and robustness in challenging tracking scenarios.
• This work opens new avenues for exploiting structural information in sparse representations for computer vision.