An Improved Method for Person Re-identification

This paper proposes a new method which combine Singular Vector Decomposition with k-reciprocal encoding for the application of person re-identification(re-ID). When we use the Euclidean distance to retrieve person, it is observed that the weight vectors in a fully connected layer are usually correlated, which makes a large impact on the retrieval result. Singular Vector Decomposition is adopted to decorrelation in this article, which has a better performance with the restraint and relaxation iteration training. Meanwhile, we add a k-reciprocal method to above result, our hypothesis is based on a gallery image is more likely to match the probe when they are in the k-reciprocal nearest neighbors. So we combine a k-reciprocal feature which is calculated by encoding its k-reciprocal nearest neighbors into a single vector under Jaccard distance and original distance as the final distance. Our method has been experimented on Market-1501 and CUHK03, it achieves a great performance, the results show that, rank-1 accuracy is improved to 82.69% and mAP is improved to 70.60% on Market-1501 for CaffeNet, while for ResNet-50, rank-1 accuracy is improved to 82.63% and mAP is improved to 73.32%.