Late Fusion in Part-based Person Re-identification

In person re-identification, the purpose is to match persons across, typically, non-overlapping cameras. This introduces challenges such as occlusion and changes in view and light- ing. In order to overcome these challenges, discriminative features are extracted and used in combination with a su- pervised metric learning algorithm. Most often, feature rep- resentations are created from the entire body, causing noisy features if certain parts are occluded. Therefore, we propose a system which applies the same learning algorithm sepa- rately on feature representations from different body parts and late fuses the outputs, to take advantage of situations in which features from certain body parts are more discrim- inative than other. By evaluation on features at three ab- straction levels, we show that the proposed system increase accuracy by up to 19.87% in the case of high-level features. In addition, we also fuse the features at different abstraction levels to further improve results. Experimental results on VIPeR and CUHK03 show similar performance to state-of- the-art with rank-1 accuracies of 52.72% and 61.50%, respec- tively, while results on the datasets PRID450S and CUHK01 show rank-1 accuracies of 78.36% and 73.40%, respectively, improvements of 11.74% and 7.76% compared to state-of- the-art.