Target Response Adaptation for Correlation Filter Tracking

Abstract

Correlation filter trackers achieve strong performance but typically use a fixed Gaussian-shaped desired response that does not adapt to changes in the target's appearance. We propose Target Response Adaptation (TRA), a method that dynamically adjusts the desired response map during tracking by learning it from the tracker's own historical output distributions. This allows the tracker to accommodate gradual appearance changes, partial occlusions, and pose variation while maintaining discriminability against background clutter. Our approach is simple to integrate into existing correlation filter frameworks and yields consistent improvements, achieving state-of-the-art results on the VOT2016 and OTB benchmarks.

Resources

Video

Citation

@inproceedings{bibi2016target,
  title     = {Target Response Adaptation for Correlation Filter Tracking},
  author    = {Bibi, Adel and M{\"{u}}ller, Matthias and Ghanem, Bernard},
  booktitle = {European Conference on Computer Vision (ECCV)},
  note      = {Spotlight},
  year      = {2016}
}