Contextual Bag-of-Words for Robust Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Appearance models are crucial for visual trackers.
• Existing models struggle with occlusion and drifting.

Purpose of the Study:

• Propose a novel contextual bag-of-words (CBOW) discriminative appearance model.
• Address challenges of drifting and occlusion in visual tracking.

Main Methods:

• Construct a compact representation using two bags-of-words from target and surrounding regions.
• Employ Bayesian inference for discriminative appearance learning.
• Introduce an adaptive updating approach with background BOWs as a sentinel to prevent drifting.
• Integrate an explicit detection method for severe occlusions.
• Implement trackers using handcrafted or deep convolutional features.

Main Results:

• Trackers demonstrate competitive performance against state-of-the-art on OTB50 and VOT2015 benchmarks.
• Outperform recent bag-of-words trackers significantly.
• Introduce the mean maximum-tracked-frame ratio (MTFR) for evaluating temporal persistence.
• Achieve superior robustness compared to other trackers based on MTFR.

Conclusions:

• The proposed CBOW model effectively handles drifting and occlusion in visual tracking.
• CBOW-based trackers show robust and competitive performance.
• The MTFR metric provides a valuable evaluation of tracker persistence.