A Self-Supervised Gait Encoding Approach for 3D Skeleton-Based Person Re-Identification with Locality Awareness

PROJECT TITLE :

A Self-Supervised Gait Encoding Approach with Locality-Awareness for 3D Skeleton Based Person Re-Identification

ABSTRACT:

The concept of person re-identification (Re-ID) using gait features within 3D skeleton sequences is a relatively new one that has several potential benefits. The currently available options rely either on manually crafted descriptors or supervised gait representation learning. In this paper, we propose a self-supervised gait encoding approach that can learn gait representations for person re-ID by utilizing unlabeled skeleton data. To be more specific, we start by building self-supervision by first learning to reconstruct unlabeled skeleton sequences backwards. This requires richer high-level semantics so that we can get better gait representations. In an effort to further improve self-supervised learning, other pretext tasks are also being investigated. We propose a locality-aware attention mechanism and a locality-aware contrastive learning scheme, both of which aim to preserve locality-awareness on intra-sequence level and inter-sequence level respectively during self-supervised learning. These mechanisms are inspired by the fact that motion's continuity endows adjacent skeletons in one skeleton sequence and temporally consecutive skeleton sequences with higher correlations (referred to as locality in 3D skeleton data). Last but not least, in order to effectively represent gait, a novel feature known as Contrastive Attention-based Gait Encodings (CAGEs) has been designed. This feature makes use of context vectors that have been learned by our locality-aware attention mechanism and contrastive learning scheme. The results of empirical tests demonstrate that our method achieves superior performance to a variety of multi-modal approaches with additional RGB or depth information, and it also significantly outperforms its skeleton-based counterparts by a margin of 15-40 percentage points in terms of Rank-1 accuracy. Our codes are available at https://github.com/Kali-Hac/Locality-Awareness-SGE .

Did you like this research project?

To get this research project Guidelines, Training and Code... Click Here