Context-Aware Correlation Filter Tracking

Abstract

Correlation filter tracking has become increasingly popular due to its computational efficiency. However, most trackers only use a local patch around the target, making them vulnerable to distractors with similar appearance. We propose a Context-Aware Correlation Filter (CA-CF) tracker that significantly improves discriminability by incorporating surrounding context regions during filter learning. We formulate the context penalty as an additional constraint in the optimization, which can be solved efficiently in the frequency domain. We also introduce a dual-regression strategy for robust scale estimation. Extensive experiments on VOT2016 and OTB-100 benchmarks demonstrate state-of-the-art performance. The work was accepted as an Oral presentation at CVPR 2017.

Resources

Video

Citation

@inproceedings{mueller2017context,
  title     = {Context-Aware Correlation Filter Tracking},
  author    = {M{\"{u}}ller, Matthias and Smith, Neil and Ghanem, Bernard},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  note      = {Oral},
  year      = {2017}
}