Dual semantic-aligned clustering for cross-domain person re-identification

Person re-identification (ReID) intends to link people with common identities captured by various cameras spread across a big, fragmented area. Due to the imbalance in the distribution of data between input and output variables and the absence of labeled information in the target domain, cross-domain person re-identification (ReID) is a high demanding assignment. This paper presents a contemporary dual semantic-aligned (DSA) clustering approach, which jointly learns latent semantic subspaces of source and target domains to group target samples. More specifically, we present an innovative subspace learning model which jointly finds intra-domain basis and inter-domain multi-granularity basis. After the subspaces are learned, target samples are correspondingly projected and clustered automatically in each subspace. The clusters allow the model to attribute imaginary labels to the target samples and then impose supervised constraints throughout the training phase. Experiment results on DukeMTMC-reID, Market-1501, and MSMT17 benchmarks validate that our scheme significantly outperformed the state-of-the-art schemes. Our approach obtains performance of 86.82%/63.06% (Duke\(\rightarrow\)Market), 76.3%/58.35% (Market\(\rightarrow\)Duke) and 52.97%/22.44% (Duke\(\rightarrow\)MSMT17) in Rank-1/mAP, correspondingly.