A Relative Positional Embedding Scheme for Transformer-Based Person Re-Identification

초록

In this letter, we propose a training scheme for a transformer-based person re-identification model using relative positional embeddings. To overcome the limitations of existing methods that rely on the visual information of an image, we define the topological and positional characteristics of a person's body structure through relative positional embeddings and uses them as an additional cue. In a set of experiment conducted for five popular person ReID benchmark datasets, the proposed scheme brings promising improvement. © 2023, Korean Institute of Communications and Information Sciences. All rights reserved.

키워드