Multi Feature Fusion at Score-Level for Appearance-based Person Re-Identification



This MATLAB-Code provides algorithms for score-level fusion with application to person re-identification.

This includes eight methods for score normalization

  • Product of likelihood-ratios
  • Logistic regression
  • FAR-based
  • Min-max
  • Z-score
  • Decimal scaling
  • Doble sigmoid
  • Tanh-estimators

and eight methods for feature weighting

  • Equal weighting
  • EER-based (equal error rate)
  • Based on Rank 1 performance
  • Based on Rank 10 performance
  • nAUC-based (normalized area under cumulative matching characteristic (CMC) curve)
  • D-prime
  • NCW (non-confidence width)
  • PROPER (pairwise optimization of projected genuine-impostor overlap)

The framework provides interfaces for evaluation on the frequently used person re-identification benchmark datasets

  • VIPeR
  • iLIDS
  • ETHZ

The framework also contains implementations of publicly available features for appearance-based person re-identification

  • Texture features: Local Binary Pattern (LBP), Maximum Response (MR8)
  • Color features: Black Value Tint (BVT) histogram, Lightness Color Opponent (Lab) histogram, weighted Hue Saturation Value (wHSV) histogram, Maximum Stable Color Regions (MSCR)
  • Combinations: Ensemble of Localized Features (ELF), Salient Dense Correspondence (SDC)

Additionally, the framework provides interfaces to combine score-level fusion with feature-level fusion (metric learning with KISSME or kernel-LFDA).


If you consider using the code provided on this page, please reference the following:

Eisenbach, M., Kolarow, A., Vorndran, A., Niebling, J., Gross, H.-M.
Evaluation of Multi Feature Fusion at Score-Level for Appearance-based Person Re-Identification
in:Proc. of Int. Joint Conf. on Neural Networks (IJCNN 2015), Killarney, Ireland, pp. 469-476, IEEE 2015


Please contact for requests.