Person Re-Identification Based on Feature Stitching

A convolutional neural networks based algorithm is proposed to extract multiple features from a single person. Further, the feature representation of a person using spliced multi-features is also proposed. Initially, the multi-branch structure is constructed using global pooling and multiple convolution; this multi-branch structure is used to offset the information loss. Subsequently, the bottleneck layer is designed to replace the classification layer in the model to reduce overfitting. In the experiment, the proposed algorithm is verified using the Market1501, CUHK03, and DukeMTMC-Reid datasets. In Market1501, the proposed algorithm achieves the first correct prediction probability (Rank1) of 95.2% and mean average precision (mAP) of 86.0%. The experimental results indicate that the proposed algorithm can extract discriminative features. Furthermore, the recognition accuracy of the proposed algorithm is significantly better than that of other advanced algorithms.