Possibilistic rank-level fusion method for person re-identification

The fusion of multiple classifiers may generate a more efficient classification than each of the individual ones. Possibility theory is particularly efficient in combining multiple information sources providing incomplete, imprecise, and conflicting knowledge. In this work, we focus on the enhancement of the person re-identification performance by combining multiple deep learning classifiers' outputs trained on different body part streams. We propose a possibilistic rank-level late fusion method that allows us to deal with imprecision and uncertainty factors that may arise in the predictions of poor classifiers. The proposed fusion method takes place in the framework of possibility theory and combines the ranking identities generated by each classifier based on their possibility distributions. This fusion method can take advantage of the complementary information given by each classifier, even the weak ones. We demonstrate the effectiveness of our proposed fusion method by presenting experimental results on two benchmark datasets (Market-1501 and DukeMTMC-reID). The obtained results show consistent accuracy improvements in comparison with state-of-the-art methods.