Person re-identification, which aims to match the same individual through different cameras, has attracted attention in recent years due to its wide application in video surveillance, and this technique is the basis for several applications suchas risk detection and behavioral judgment.
In order to solve the local optimum problem of person re-identification, we have studied a novel neural network learning or deep feature learning method for mixed distance maximization based on full-triple relation.
Unlike traditional triplet loss, the objective function is designed to simultaneously consider the distances between triplets as well as the distances within triplet.
First, we defined a local objective function to maximize the inter-distance for person re-identification. Maximization of this function leads to further reduction of the distance between matching pairs and further increase of the distance between mismatching pairs.
Second, we also developed a global objective function to take into account the distance between triplets. The global objective function was defined by distances between triplets. This function allows different triplets to be closer in the feature space.
Finally, to optimize the information of triplets, we defined a main objective function with the mixed distance based on the so-called full-triple relation combiningthe two distances.
We have verified the excellent performance of the proposed method on several public benchmark datasets and conducted experimental analysis on the public datasets, i.e., VIPeR and ETHZ.
The results showed that the proposed deep learning framework with mixed distance maximization based on full-triple relation can be applied to person re-identification and achieve the desired performance.
The results were published in the "Multimedia Tools and Applications"(2024) under the title of "Deep Learning with full-triple relational person re-identification"(https://doi.org/10.1007/s11042-024-20550-y).
