Re-ranking person re-identification using distance aggregation of k-nearest neighbors hierarchical tree

Abstract

Listen

Translate article

Person re-identification is a challenging task due to the critical factors like illumination, occlusion, pose variation, view-points, low resolution, inter/intra-class variations, etc. Re-identification mainly treated as the target retrieval process, which can be improved by multi-query and re-ranking. Due to the high computational cost and consideration of fixed length gallery size, existing re-ranking approaches are not feasible for real-time re-identification applications, specifically with the variable-length gallery. We have proposed a fast yet effective re-ranking approach that utilize the pre-computed pair-wise distance used for initial ranking. We have integrated the advantages of nearest neighbors, hierarchical tree, and multi-query for re-ranking. We hypothesize that the hierarchy of k-nearest neighbors of an image leads to more positive matches in the child layer. Hence the aggregated distance decreases for true match and increases for false match images of the initial rank. We have structured k-nearest neighbor’s hierarchical tree and calculated the aggregated distance. Hierarchical nearest neighbors are treated as multi-query under distance aggregation. Final re-ranked distance is computed as the weighted sum of aggregated and actual distance. Our proposed re-ranking approach is computationally efficient, feasible for real-time applications, unsupervised, and completely automatic. The effectiveness of our proposed method (Source code isavailable upon request) has been verified by various experiments on MARS, Market-1501, DukeMTMC, and CUHK03 datasets.