Multi-granularity graph pooling for video-based person re-identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video-based person re-identification (ReID) is crucial for surveillance and security.
• Graph Neural Networks (GNNs) are used to aggregate spatio-temporal features in video ReID.
• Existing GNN models often neglect graph topology and node importance during feature aggregation.

Purpose of the Study:

• To propose a novel Graph Pooling Network (GPNet) for improved video-based person re-identification.
• To develop a multi-granular graph representation learning method for video retrieval.
• To address limitations of mean/max pooling in existing graph-based ReID models.

Main Methods:

• Constructing a multi-granular graph using backbone-learned node features.
• Employing multiple graph convolutional layers for spatial and temporal feature aggregation.
• Introducing a Multi-Head Full Attention Graph Pooling (MHFAPool) layer for graph downsampling, integrating node clustering and selection with attention mechanisms.

Main Results:

• GPNet achieves competitive results on widely-used person re-identification datasets (MARS, DukeMTMC-VideoReID, iLIDS-VID, PRID-2011).
• The proposed MHFAPool effectively learns graph representations by considering node importance and graph topology.
• The multi-granular graph representation captures richer information for video retrieval.

Conclusions:

• GPNet offers a superior approach to video-based person re-identification by enhancing graph representation learning.
• The novel MHFAPool layer is effective in downsampling graphs while preserving crucial topological and feature information.
• The proposed method demonstrates strong performance across multiple benchmark datasets, validating its efficacy.